#### PRF-Ped: Multi-Scale Pedestrian Detector with Prior-Based Receptive Field

Yuzhi Tan, Hongxun Yao, Haoran Li, Xiusheng Lu, Haozhe Xie
Track 3: Computer Vision Robotics and Intelligent Systems
Thu 14 Jan 2021 at 12:00 in session PS T3.7

Auto-TLDR; Bidirectional Feature Enhancement Module for Multi-Scale Pedestrian Detection

Multi-scale feature representation is a common strategy to handle the scale variation in pedestrian detection. Existing methods simply utilize the convolutional pyramidal features for multi-scale representation. However, they rarely pay attention to the differences among different feature scales and extract multi-scale features from a single feature map, which may make the detectors sensitive to scale-variance in multi-scale pedestrian detection. In this paper, we introduce a bidirectional feature enhancement module (BFEM) to augment the semantic information of low-level features and the localization information of high-level features. In addition, we propose a prior-based receptive field block (PRFB) for multi-scale pedestrian feature extraction, where the receptive field is closer to the aspect ratio of the pedestrian target. Consequently, it is less affected by the surrounding background when extracting features. Experimental results indicate that the proposed method outperform the state-of-the-art methods on the CityPersons and Caltech datasets.

## Similar papers

#### Learning a Dynamic High-Resolution Network for Multi-Scale Pedestrian Detection

Mengyuan Ding, Shanshan Zhang, Jian Yang
Track 3: Computer Vision Robotics and Intelligent Systems
Thu 14 Jan 2021 at 16:00 in session PS T3.9

Auto-TLDR; Learningable Dynamic HRNet for Pedestrian Detection

Pedestrian detection is a canonical instance of object detection in computer vision. In practice, scale variation is one of the key challenges, resulting in unbalanced performance across different scales. Recently, the High-Resolution Network (HRNet) has become popular because high-resolution feature representations are more friendly to small objects. However, when we apply HRNet for pedestrian detection, we observe that it improves for small pedestrians on one hand, but hurts the performance for larger ones on the other hand. To overcome this problem, we propose a learnable Dynamic HRNet (DHRNet) aiming to generate different network paths adaptive to different scales. Specifically, we construct a parallel multi-branch architecture and add a soft conditional gate module allowing for dynamic feature fusion. Both branches share all the same parameters except the soft gate module. Experimental results on CityPersons and Caltech benchmarks indicate that our proposed dynamic HRNet is more capable of dealing with pedestrians of various scales, and thus improves the performance across different scales consistently.

#### Nighttime Pedestrian Detection Based on Feature Attention and Transformation

Gang Li, Shanshan Zhang, Jian Yang
Track 3: Computer Vision Robotics and Intelligent Systems
Thu 14 Jan 2021 at 16:00 in session PS T3.9

Auto-TLDR; FAM and FTM: Enhanced Feature Attention Module and Feature Transformation Module for nighttime pedestrian detection

Pedestrian detection at nighttime is an important yet challenging task, which is fundamental for many practical applications, e.g. autonomous driving, video surveillance. To address this problem, in this work we start with some analysis, from which we find that the nighttime features have much more noise than that of daytime, resulting in low discrimination ability. Besides, we also observe some pedestrian examples are under adverse illumination conditions, and they can hardly provide sufficient information for accurate detection. Based on these findings, we propose the Feature Attention Module (FAM) and Feature Transformation Module (FTM) to enhance nighttime features. In FAM, guided by progressive segmentation supervision, hierarchical feature attention is produced to enhance multi-level features. On the other hand, FTM is introduced to enforce features from adverse illumination to approach that from better illumination. Based on feature attention and transformation (FAT) mechanism, a two-stage detector called FATNet is constructed for nighttime pedestrian detection. We conduct extensive experiments on nighttime datasets of EuroCity Persons (Night) and NightOwls to demonstrate the effectiveness of our method. On both two datasets, our method achieves significant improvements to the baseline and also outperforms state-of-the-art detectors.

#### Bidirectional Matrix Feature Pyramid Network for Object Detection

Wei Xu, Yi Gan, Jianbo Su
Track 3: Computer Vision Robotics and Intelligent Systems
Fri 15 Jan 2021 at 15:00 in session PS T3.10

Auto-TLDR; BMFPN: Bidirectional Matrix Feature Pyramid Network for Object Detection

Feature pyramids are widely used to improve scale invariance for object detection. Most methods just map the objects to feature maps with relevant square receptive fields, but rarely pay attention to the aspect ratio variation, which is also an important property of object instances. It will lead to a poor match between rectangular objects and assigned features with square receptive fields, thus preventing from accurate recognition and location. Besides, the information propagation among feature layers is sparse, namely, each feature in the pyramid may mainly or only contain single-level information, which is not representative enough for classification and localization sub-tasks. In this paper, Bidirectional Matrix Feature Pyramid Network (BMFPN) is proposed to address these issues. It consists of three modules: Diagonal Layer Generation Module (DLGM), Top-down Module (TDM) and Bottom-up Module (BUM). First, multi-level features extracted by backbone are fed into DLGM to produce the base features. Then these base features are utilized to construct the final feature pyramid through TDM and BUM in series. The receptive fields of the designed feature layers in BMFPN have various scales and aspect ratios. Objects can be correctly assigned to appropriate and representative feature maps with relevant receptive fields depending on its scale and aspect ratio properties. Moreover, TDM and BUM form bidirectional and reticular information flow, which effectively fuses multi level information in top-down and bottom-up manner respectively. To evaluate the effectiveness of our proposed architecture, an end-toend anchor-free detector is designed and trained by integrating BMFPN into FCOS. And the center ness branch in FCOS is modified with our Gaussian center-ness branch (GCB), which brings another slight improvement. Without bells and whistles, our method gains +3.3%, +2.4% and +2.6% AP on MS COCO dataset from baselines with ResNet-50, ResNet-101 and ResNeXt-101 backbones, respectively.

#### MagnifierNet: Learning Efficient Small-Scale Pedestrian Detector towards Multiple Dense Regions

Qi Cheng, Mingqin Chen, Yingjie Wu, Fei Chen, Shiping Lin
Track 3: Computer Vision Robotics and Intelligent Systems
Wed 13 Jan 2021 at 12:00 in session PS T3.4

Auto-TLDR; MagnifierNet: A Simple but Effective Small-Scale Pedestrian Detection Towards Multiple Dense Regions

Despite the success of pedestrian detection, there is still a significant gap in the performance of the detection of pedestrians at different scales. Detecting small-scale pedestrians is extremely challenging due to the low resolution of their convolution features which is essential for downstream classifiers. To address this issue, we observed pedestrian datasets and found that pedestrians often gather together in crowded public places. Then we propose MagnifierNet, a simple but effective small-scale pedestrian detector towards multiple dense regions. MagnifierNet uses our proposed sweep-line based grouping algorithm to find dense regions based on the number of pedestrians in the grouped region. And we adopt a new definition of small-scale pedestrians through grid search and KL-divergence. Besides, our grouping method can also be used as a new strategy for pedestrian data augmentation. The ablation study demonstrates that MagnifierNet improves the representation of small-scale pedestrians. We validate the effectiveness of MagnifierNet on CityPersons and KITTI datasets. Experimental results show that MagnifierNet achieves the best small-scale pedestrian detection performance on CityPersons benchmark without any external data, and also achieves competitive performance for detecting small-scale pedestrians on KITTI dataset without bells and whistles.

#### Construction Worker Hardhat-Wearing Detection Based on an Improved BiFPN

Chenyang Zhang, Zhiqiang Tian, Jingyi Song, Yaoyue Zheng, Bo Xu
Track 3: Computer Vision Robotics and Intelligent Systems
Tue 12 Jan 2021 at 15:00 in session PS T3.1

Auto-TLDR; A One-Stage Object Detection Method for Hardhat-Wearing in Construction Site

Work in the construction site is considered to be one of the occupations with the highest safety risk factor. Therefore, safety plays an important role in construction site. One of the most fundamental safety rules in construction site is to wear a hardhat. To strengthen the safety of the construction site, most of the current methods use multi-stage method for hardhat-wearing detection. These methods have limitations in terms of adaptability and generalizability. In this paper, we propose a one-stage object detection method based on convolutional neural network. We present a multi-scale strategy that selects the high-resolution feature maps of DarkNet-53 to effectively identify small-scale hardhats. In addition, we propose an improved weighted bi-directional feature pyramid network (BiFPN), which could fuse more semantic features from more scales. The proposed method can not only detect hardhat-wearing, but also identify the color of the hardhat. Experimental results show that the proposed method achieves a mAP of 87.04%, which outperforms several state-of-the-art methods on a public dataset.

#### CenterRepp: Predict Central Representative Point Set's Distribution for Detection

Yulin He, Limeng Zhang, Wei Chen, Xin Luo, Chen Li, Xiaogang Jia
Track 3: Computer Vision Robotics and Intelligent Systems
Thu 14 Jan 2021 at 16:00 in session PS T3.9

Auto-TLDR; CRPDet: CenterRepp Detector for Object Detection

Object detection has long been an important issue in the discipline of scene understanding. Existing researches mainly focus on the object itself, ignoring its surrounding environment. In fact, the surrounding environment provides abundant information to help detectors classify and locate objects. This paper proposes CRPDet, viz. CenterRepp Detector, a framework for object detection. The main function of CRPDet is accomplished by the CenterRepp module, which takes into account the surrounding environment by predicting the distribution of the central representative points. CenterRepp converts labeled object frames into the mean and standard variance of the sampling points’ distribution. This helps increase the receptive field of objects, breaking the limitation of object frames. CenterRepp defines a position-fixed center point with significant weights, avoiding to sample all points in the surroundings. Experiments on the COCO test-dev detection benchmark demonstrates that our proposed CRPDet has comparable performance with state-of-the-art detectors, achieving 39.4 mAP with 51 FPS tested under single size input.

#### Mutual-Supervised Feature Modulation Network for Occluded Pedestrian Detection

Ye He, Chao Zhu, Xu-Cheng Yin
Track 3: Computer Vision Robotics and Intelligent Systems
Wed 13 Jan 2021 at 14:00 in session OS T3.2

Auto-TLDR; A Mutual-Supervised Feature Modulation Network for Occluded Pedestrian Detection

State-of-the-art pedestrian detectors have achieved significant progress on non-occluded pedestrians, yet they are still struggling under heavy occlusions. The recent occlusion handling strategy of popular two-stage approaches is to build a two-branch architecture with the help of additional visible body annotations. Nonetheless, these methods still have some weaknesses. Either the two branches are trained independently with only score-level fusion, which cannot guarantee the detectors to learn robust enough pedestrian features. Or the attention mechanisms are exploited to only emphasize on the visible body features. However, the visible body features of heavily occluded pedestrians are concentrated on a relatively small area, which will easily cause missing detections. To address the above issues, we propose in this paper a novel Mutual-Supervised Feature Modulation (MSFM) network, to better handle occluded pedestrian detection. The key MSFM module in our network calculates the similarity loss of full body boxes and visible body boxes corresponding to the same pedestrian, so that the full-body detector could learn more complete and robust pedestrian features with the assist of contextual features from the occluding parts. To facilitate the MSFM module, we also propose a novel two-branch architecture, consisting of a standard full body detection branch and an extra visible body classification branch. These two branches are trained in a mutual-supervised way with full body annotations and visible body annotations, respectively. To verify the effectiveness of our proposed method, extensive experiments are conducted on two challenging pedestrian datasets: Caltech and CityPersons, and our approach achieves superior performances compared to other state-of-the-art methods on both datasets, especially in heavy occlusion cases.

#### SFPN: Semantic Feature Pyramid Network for Object Detection

Yi Gan, Wei Xu, Jianbo Su
Track 1: Artificial Intelligence, Machine Learning for Pattern Analysis
Fri 15 Jan 2021 at 15:00 in session PS T1.13

Auto-TLDR; SFPN: Semantic Feature Pyramid Network to Address Information Dilution Issue in FPN

Feature Pyramid Network(FPN) employs a top-down path to enhance low level feature by utilizing high level feature.However, further improvement of detector is greatly hindered by the inner defect of FPN. The dilution issue in FPN is analyzed in this paper, and a new architecture named Semantic Feature Pyramid Network(SFPN) is introduced to address the information imbalance problem caused by information dilution. The proposed method consists of two simple and effective components: Semantic Pyramid Module(SPM) and Semantic Feature Fusion Module(SFFM). To compensate for the weaknesses of FPN, the semantic segmentation result is utilized as an extra information source in our architecture.By constructing a semantic pyramid based on the segmentation result and fusing it with FPN, feature maps at each level can obtain the necessary information without suffering from the dilution issue. The proposed architecture could be applied on many detectors, and non-negligible improvement could be achieved. Although this method is designed for object detection, other tasks such as instance segmentation can also largely benefit from it. The proposed method brings Faster R-CNN and Mask R-CNN with ResNet-50 as backbone both 1.8 AP improvements respectively. Furthermore, SFPN improves Cascade R-CNN with backbone ResNet-101 from 42.4 AP to 43.5 AP.

#### Efficient-Receptive Field Block with Group Spatial Attention Mechanism for Object Detection

Jiacheng Zhang, Zhicheng Zhao, Fei Su
Track 3: Computer Vision Robotics and Intelligent Systems
Thu 14 Jan 2021 at 12:00 in session PS T3.7

Auto-TLDR; E-RFB: Efficient-Receptive Field Block for Deep Neural Network for Object Detection

Object detection has been paid rising attention in computer vision field. Convolutional Neural Networks (CNNs) extract high-level semantic features of images, which directly determine the performance of object detection. As a common solution, embedding integration modules into CNNs can enrich extracted features and thereby improve the performance. However, the instability and inconsistency of internal multiple branches exist in these modules. To address this problem, we propose a novel multibranch module called Efficient-Receptive Field Block (E-RFB), in which multiple levels of features are combined for network optimization. Specifically, by downsampling and increasing depth, the E-RFB provides sufficient RF. Second, in order to eliminate the inconsistency across different branches, a novel spatial attention mechanism, namely, Group Spatial Attention Module (GSAM) is proposed. The GSAM gradually narrows a feature map by channel grouping; thus it encodes the information between spatial and channel dimensions into the final attention heat map. Third, the proposed module can be easily joined in various CNNs to enhance feature representation as a plug-and-play component. With SSD-style detectors, our method halves the parameters of the original detection head and achieves high accuracy on the PASCAL VOC and MS COCO datasets. Moreover, the proposed method achieves superior performance compared with state-of-the-art methods based on similar framework.

#### DualBox: Generating BBox Pair with Strong Correspondence Via Occlusion Pattern Clustering and Proposal Refinement

Zheng Ge, Chuyu Hu, Xin Huang, Baiqiao Qiu, Osamu Yoshie
Track 3: Computer Vision Robotics and Intelligent Systems
Wed 13 Jan 2021 at 12:00 in session PS T3.4

Auto-TLDR; R2NMS: Combining Full and Visible Body Bounding Box for Dense Pedestrian Detection

Despite the rapid development of pedestrian detection, the problem of dense pedestrian detection is still unsolved, especially the upper limit of Recall caused by Non-Maximum-Suppression (NMS). Out of this reason, R2NMS is proposed to simultaneously detect full and visible body bounding boxes, by replacing the full body BBoxes with less occluded visible body BBoxes in the NMS algorithm, achieving a higher recall. However, the P-RPN and P-RCNN modules proposed in R2NMS for simultaneous high quality full and visible body prediction require non-trivial positive/negative assigning strategies for anchor BBoxes. To simplify the prerequisites and improve the utility of R2NMS, we incorporate clustering analysis into the learning of visible body proposals from full body proposals. Furthermore, to reduce the computation complexity caused by the large number of potential visible body proposals, we introduce a novel occlusion pattern prediction branch on top of the R-CNN module (i.e. F-RCNN) to select the best matched visible proposals for each full body proposals and then feed them into another R-CNN module (i.e. V-RCNN). Incorporated with R2NMS, our DualBox model can achieve competitive performance while only requires few hyper-parameters. We validate the effectiveness of the proposed approach on the CrowdHuman and CityPersons datasets. Experimental results show that our approach achieves promising performance for detecting both non-occluded and occluded pedestrians, especially heavily occluded ones.

#### Convolutional STN for Weakly Supervised Object Localization

Akhil Meethal, Marco Pedersoli, Soufiane Belharbi, Eric Granger
Track 3: Computer Vision Robotics and Intelligent Systems
Fri 15 Jan 2021 at 13:00 in session OS T3.5

Auto-TLDR; Spatial Localization for Weakly Supervised Object Localization

Weakly-supervised object localization is a challenging task in which the object of interest should be localized while learning its appearance. State-of-the-art methods recycle the architecture of a standard CNN by using the activation maps of the last layer for localizing the object. While this approach is simple and works relatively well, object localization relies on different features than classification, thus, a specialized localization mechanism is required during training to improve performance. In this paper, we propose a convolutional, multi-scale spatial localization network that provides accurate localization for the object of interest. Experimental results on CUB-200-2011 and ImageNet datasets show competitive performance of our proposed approach on Weakly supervised localization.

#### Triplet-Path Dilated Network for Detection and Segmentation of General Pathological Images

Jiaqi Luo, Zhicheng Zhao, Fei Su, Limei Guo
Track 5: Image and Signal Processing
Fri 15 Jan 2021 at 13:00 in session OS T5.5

Auto-TLDR; Triplet-path Network for One-Stage Object Detection and Segmentation in Pathological Images

Deep learning has been widely applied in the field of medical image processing. However, compared with flourishing visual tasks in natural images, the progress achieved in pathological images is not remarkable, and detection and segmentation, which are among basic tasks of computer vision, are regarded as two independent tasks. In this paper, we make full use of existing datasets and construct a triplet-path network using dilated convolutions to cooperatively accomplish one-stage object detection and nuclei segmentation for general pathological images. First, in order to meet the requirement of detection and segmentation, a novel structure called triplet feature generation (TFG) is designed to extract high-resolution and multiscale features, where features from different layers can be properly integrated. Second, considering that pathological datasets are usually small, a location-aware and partially truncated loss function is proposed to improve the classification accuracy of datasets with few images and widely varying targets. We compare the performance of both object detection and instance segmentation with state-of-the-art methods. Experimental results demonstrate the effectiveness and efficiency of the proposed network on two datasets collected from multiple organs.

#### Cascade Saliency Attention Network for Object Detection in Remote Sensing Images

Dayang Yu, Rong Zhang, Shan Qin
Track 3: Computer Vision Robotics and Intelligent Systems
Thu 14 Jan 2021 at 14:00 in session PS T3.8

Auto-TLDR; Cascade Saliency Attention Network for Object Detection in Remote Sensing Images

Object detection in remote sensing images is a challenging task due to objects in the bird-view perspective appearing with arbitrary orientations. Though considerable progress has been made, there still exist challenges with the interference from complex backgrounds, dense arrangement, and large-scale variations. In this paper, we propose an oriented detector named Cascade Saliency Attention Network (CSAN), designed for comprehensively suppressing interference in remote sensing images. Specifically, we first combine context and pixel attention on feature maps to enhance saliency of objects for suppressing interference from backgrounds. Then, in cascade network, we apply instance segmentation on ROI to increase saliency of the central object, thus preventing object features from mutual interference in dense arrangement. Additionally, to alleviate large-scale variations, we devise a multi-scale merge module during FPN merging process to learn richer scale representations. Experimental results on DOTA and HRSC2016 datasets outperform other state-of-the-art object detection methods and verify the effectiveness of our method.

#### Scene Text Detection with Selected Anchors

Anna Zhu, Hang Du, Shengwu Xiong
Track 4: Document and Media Analysis
Thu 14 Jan 2021 at 12:00 in session PS T4.3

Auto-TLDR; AS-RPN: Anchor Selection-based Region Proposal Network for Scene Text Detection

Object proposal technique with dense anchoring scheme for scene text detection were applied frequently to achieve high recall. It results in the significant improvement in accuracy but waste of computational searching, regression and classification. In this paper, we propose an anchor selection-based region proposal network (AS-RPN) using effective selected anchors instead of dense anchors to extract text proposals. The center, scales, aspect ratios and orientations of anchors are learnable instead of fixing, which leads to high recall and greatly reduced numbers of anchors. By replacing the anchor-based RPN in Faster RCNN, the AS-RPN-based Faster RCNN can achieve comparable performance with previous state-of-the-art text detecting approaches on standard benchmarks, including COCO-Text, ICDAR2013, ICDAR2015 and MSRA-TD500 when using single-scale and single model (ResNet50) testing only.

#### EDD-Net: An Efficient Defect Detection Network

Tianyu Guo, Linlin Zhang, Runwei Ding, Ge Yang
Track 3: Computer Vision Robotics and Intelligent Systems
Thu 14 Jan 2021 at 16:00 in session PS T3.9

Auto-TLDR; EfficientNet: Efficient Network for Mobile Phone Surface defect Detection

As the most commonly used communication tool, the mobile phone has become an indispensable part of our daily life. The surface of the mobile phone as the main window of human-phone interaction directly affects the user experience. It is necessary to detect surface defects on the production line in order to ensure the high quality of the mobile phone. However, the existing mobile phone surface defect detection is mainly done manually, and currently there are few automatic defect detection methods to replace human eyes. How to quickly and accurately detect the surface defects of mobile phone is an urgent problem to be solved. Hence, an efficient defect detection network (EDD-Net) is proposed. Firstly, EfficientNet is used as the backbone network. Then, according to the small-scale of mobile phone surface defects, a feature pyramid module named GCSA-BiFPN is proposed to obtain more discriminative features. Finally, the box/class prediction network is used to achieve effective defect detection. We also build a mobile phone surface oil stain defect (MPSOSD) dataset to alleviate the lack of dataset in this field. The performance on the relevant datasets shows that the network we proposed is effective and has practical significance for industrial production.

#### ScarfNet: Multi-Scale Features with Deeply Fused and Redistributed Semantics for Enhanced Object Detection

Jin Hyeok Yoo, Dongsuk Kum, Jun Won Choi
Track 3: Computer Vision Robotics and Intelligent Systems
Fri 15 Jan 2021 at 15:00 in session PS T3.10

Auto-TLDR; Semantic Fusion of Multi-scale Feature Maps for Object Detection

Convolutional neural networks (CNNs) have led us to achieve significant progress in object detection research. To detect objects of various sizes, object detectors often exploit the hierarchy of the multiscale feature maps called {\it feature pyramids}, which are readily obtained by the CNN architecture. However, the performance of these object detectors is limited because the bottom-level feature maps, which experience fewer convolutional layers, lack the semantic information needed to capture the characteristics of the small objects. To address such problems, various methods have been proposed to increase the depth for the bottom-level features used for object detection. While most approaches are based on the generation of additional features through the top-down pathway with lateral connections, our approach directly fuses multi-scale feature maps using bidirectional long short-term memory (biLSTM) in an effort to leverage the gating functions and parameter-sharing in generating deeply fused semantics. The resulting semantic information is redistributed to the individual pyramidal feature at each scale through the channel-wise attention model. We integrate our semantic combining and attentive redistribution feature network (ScarfNet) with the baseline object detectors, i.e., Faster R-CNN, single-shot multibox detector (SSD), and RetinaNet. Experimental results show that our method offers a significant performance gain over the baseline detectors and outperforms the competing multiscale fusion methods in the PASCAL VOC and COCO detection benchmarks.

#### HPERL: 3D Human Pose Estimastion from RGB and LiDAR

Michael Fürst, Shriya T.P. Gupta, René Schuster, Oliver Wasenmüler, Didier Stricker
Track 3: Computer Vision Robotics and Intelligent Systems
Tue 12 Jan 2021 at 17:00 in session PS T3.3

Auto-TLDR; 3D Human Pose Estimation Using RGB and LiDAR Using Weakly-Supervised Approach

In-the-wild human pose estimation has a huge potential for various fields, ranging from animation and action recognition to intention recognition and prediction for autonomous driving. The current state-of-the-art is focused only on RGB and RGB-D approaches for predicting the 3D human pose. However, not using precise LiDAR depth information limits the performance and leads to very inaccurate absolute pose estimation. With LiDAR sensors becoming more affordable and common on robots and autonomous vehicle setups, we propose an end-to-end architecture using RGB and LiDAR to predict the absolute 3D human pose with unprecedented precision. Additionally, we introduce a weakly-supervised approach to generate 3D predictions using 2D pose annotations from PedX. This allows for many new opportunities in the field of 3D human pose estimation.

#### Small Object Detection by Generative and Discriminative Learning

Yi Gu, Jie Li, Chentao Wu, Weijia Jia, Jianping Chen
Track 3: Computer Vision Robotics and Intelligent Systems
Wed 13 Jan 2021 at 12:00 in session PS T3.4

Auto-TLDR; Generative and Discriminative Learning for Small Object Detection

With the development of deep convolutional neural networks (CNNs), the object detection accuracy has been greatly improved. But the performance of small object detection is still far from satisfactory, mainly because small objects are so tiny that the information contained in the feature map is limited. Existing methods focus on improving classification accuracy but still suffer from the limitation of bounding box prediction. To solve this issue, we propose a detection framework by generative and discriminative learning. First, a reconstruction generator network is designed to reconstruct the mapping from low frequency to high frequency for anchor box prediction. Then, a detector module extracts the regions of interest (ROIs) from generated results and implements a RoI-Head to predict object category and refine bounding box. In order to guide the reconstructed image related to the corresponding one, a discriminator module is adopted to tell from the generated result and the original image. Extensive evaluations on the challenging MS-COCO dataset demonstrate that our model outperforms most state-of-the-art models in detecting small objects, especially the reconstruction module improves the average precision for small object (APs) by 7.7%.

#### Forground-Guided Vehicle Perception Framework

Kun Tian, Tong Zhou, Shiming Xiang, Chunhong Pan
Track 3: Computer Vision Robotics and Intelligent Systems
Fri 15 Jan 2021 at 15:00 in session PS T3.10

Auto-TLDR; A foreground segmentation branch for vehicle detection

As the basis of advanced visual tasks such as vehicle tracking and traffic flow analysis, vehicle detection needs to accurately predict the position and category of vehicle objects. In the past decade, deep learning based methods have made great progress. However, we also notice that some existing cases are not studied thoroughly. First, false positive on the background regions is one of the critical problems. Second, most of the previous approaches only optimize a single vehicle detection model, ignoring the relationship between different visual perception tasks. In response to the above two findings, we introduce a foreground segmentation branch for the first time, which can predict the pixel level of vehicles in advance. Furthermore, two attention modules are designed to guide the work of the detection branch. The proposed method can be easily grafted into the one-stage and two-stage detection framework. We evaluate the effectiveness of our model on LSVH, a dataset with large variations in vehicle scales, and achieve the state-of-the-art detection accuracy.

#### Efficient High-Resolution High-Level-Semantic Representation Learning for Human Pose Estimation

Hong Liu, Lisi Guan
Track 3: Computer Vision Robotics and Intelligent Systems
Fri 15 Jan 2021 at 15:00 in session PS T3.10

Auto-TLDR; Spatial enhanced separated temporal spatial convolutional neural network

Temporal-spatial information, as the most key issue for human action recognition, has been explored by lots of means, such as 3D convolution network (3DCNN) based or 3DCNN decomposing based approaches. Though the latter can be seen as a trade-off for overcoming the shortage caused by the former for reducing the computation cost and saving parameters, information imbalance of videos between spatial and temporal information is still not been well excavated. To tackle this problem, spatial enhanced separated temporal spatial convolutional neural network (SESTSN) is proposed in this paper, which can easily outperform 3DCNN based and 3DCNN decomposing based methods with fewer parameters. What's more, to further reduce parameter and computation cost, we adopt depth-wise convolution to the proposed SESTSN and propose the channel separated spatial enhanced separated temporal spatial convolutional neural network (CSESTSN). Experiments show that the proposed CSESTSN contains considerably fewer parameters involving much lower computation cost, while it achieves comparable performance to 3D convolution-based methods. Our method outperforms state-of-the-art methods on two challenging datasets, namely NTU RGB+D dataset and Northwestern-UCLA dataset, which verifies the effectiveness of our method.

#### Detecting Objects with High Object Region Percentage

Fen Fang, Qianli Xu, Liyuan Li, Ying Gu, Joo-Hwee Lim
Track 3: Computer Vision Robotics and Intelligent Systems
Thu 14 Jan 2021 at 14:00 in session PS T3.8

Auto-TLDR; Faster R-CNN for High-ORP Object Detection

Object shape is a subtle but important factor for object detection. It has been observed that the object-region-percentage (ORP) can be utilized to improve detection accuracy for elongated objects, which have much lower ORPs than other types of objects. In this paper, we propose an approach to improve the detection performance for objects whose ORPs are relatively higher.To address the problem of high-ORP object detection, we propose a method consisting of three steps. First, we adjust the ground truth bounding boxes of high-ORP objects to an optimal range. Second, we train an object detector, Faster R-CNN, based on adjusted bounding boxes to achieve high recall. Finally, we train a DCNN to learn the adjustment ratios towards four directions and adjust detected bounding boxes of objects to get better localization for higher precision. We evaluate the effectiveness of our method on 12 high-ORP objects in COCO and 8 objects in a proprietary gearbox dataset. The experimental results show that our method can achieve state-of-the-art performance on these objects while costing less resources in training and inference stages.

#### Encoder-Decoder Based Convolutional Neural Networks with Multi-Scale-Aware Modules for Crowd Counting

Pongpisit Thanasutives, Ken-Ichi Fukui, Masayuki Numao, Boonserm Kijsirikul
Track 2: Biometrics, Human Analysis and Behavior Understanding
Wed 13 Jan 2021 at 12:00 in session PS T2.2

Auto-TLDR; M-SFANet and M-SegNet for Crowd Counting Using Multi-Scale Fusion Networks

In this paper, we proposed two modified neural networks based on dual path multi-scale fusion networks (SFANet) and SegNet for accurate and efficient crowd counting. Inspired by SFANet, the first model, which is named M-SFANet, is attached with atrous spatial pyramid pooling (ASPP) and context-aware module (CAN). The encoder of M-SFANet is enhanced with ASPP containing parallel atrous convolutional layers with different sampling rates and hence able to extract multi-scale features of the target object and incorporate larger context. To further deal with scale variation throughout an input image, we leverage the CAN module which adaptively encodes the scales of the contextual information. The combination yields an effective model for counting in both dense and sparse crowd scenes. Based on the SFANet decoder structure, M-SFANet's decoder has dual paths, for density map and attention map generation. The second model is called M-SegNet, which is produced by replacing the bilinear upsampling in SFANet with max unpooling that is used in SegNet. This change provides a faster model while providing competitive counting performance. Designed for high-speed surveillance applications, M-SegNet has no additional multi-scale-aware module in order to not increase the complexity. Both models are encoder-decoder based architectures and are end-to-end trainable. We conduct extensive experiments on five crowd counting datasets and one vehicle counting dataset to show that these modifications yield algorithms that could improve state-of-the-art crowd counting methods.

#### Temporal Feature Enhancement Network with External Memory for Object Detection in Surveillance Video

Masato Fujitake, Akihiro Sugimoto
Track 5: Image and Signal Processing
Thu 14 Jan 2021 at 12:00 in session PS T5.5

Auto-TLDR; Temporal Attention Based External Memory Network for Surveillance Object Detection

Video object detection is challenging and essential in practical applications, such as surveillance cameras for traffic control and public security. Unlike the video in natural scenes, the surveillance video tends to contain dense, and small objects (typically vehicles) in their appearances. Therefore, existing methods for surveillance object detection utilize still-image object detection approaches with rich feature extractors at the expense of their run-time speeds. The run-time speed, however, becomes essential when the video is being streamed. In this paper, we exploit temporal information in videos to enrich the feature maps, proposing the first temporal attention based external memory network for the live stream of video. Extensive experiments on real-world traffic surveillance benchmarks demonstrate the real-time performance of the proposed model while keeping comparable accuracy with state-of-the-art.

#### Spatial-Related and Scale-Aware Network for Crowd Counting

Lei Li, Yuan Dong, Hongliang Bai
Track 1: Artificial Intelligence, Machine Learning for Pattern Analysis
Wed 13 Jan 2021 at 14:00 in session PS T1.5

Auto-TLDR; Spatial Attention for Crowd Counting

Crowd counting aims to estimate the number of people in images. Although promising progresses have been made with the prevalence of deep Convolutional Neural Networks, there still remains a challenging task due to cluttered backgrounds and varying scales of people within an image. In this paper, we propose a learnable spatial attention module which can get the spatial relations to diminish the negative impact of backgrounds. Besides, a dense hybrid dilated convolution module is also brought up to preserve information derived from varied scales. With these two modules, our network can deal with the problem caused by scale variance and background interference. To demonstrate the effectiveness of our method, we compare it with state-of-the-art algorithms on three representative crowd counting benchmarks (ShanghaiTech UCF-QNRF,UCF_CC_50). Experimental results show that our proposed network can achieve significant improvements on all the three datasets.

#### Tiny Object Detection in Aerial Images

Jinwang Wang, Wen Yang, Haowen Guo, Ruixiang Zhang, Gui-Song Xia
Track 3: Computer Vision Robotics and Intelligent Systems
Fri 15 Jan 2021 at 13:00 in session OS T3.5

Auto-TLDR; Tiny Object Detection in Aerial Images Using Multiple Center Points Based Learning Network

Object detection in Earth Vision has achieved great progress in recent years. However, tiny object detection in aerial images remains a very challenging problem since the tiny objects contain a small number of pixels and are easily confused with the background. To advance tiny object detection research in aerial images, we present a new dataset for Tiny Object Detection in Aerial Images (AI-TOD). Specifically, AI-TOD comes with 700,621 object instances for eight categories across 28,036 aerial images. Compared to existing object detection datasets in aerial images, the mean size of objects in AI-TOD is about 12.8 pixels, which is much smaller than others. To build a benchmark for tiny object detection in aerial images, we evaluate the state-of-the-art object detectors on our AI-TOD dataset. Experimental results show that direct application of these approaches on AI-TOD produces suboptimal object detection results, thus new specialized detectors for tiny object detection need to be designed. Therefore, we propose a multiple center points based learning network (M-CenterNet) to improve the localization performance of tiny object detection, and experimental results show the significant performance gain over the competitors.

#### SyNet: An Ensemble Network for Object Detection in UAV Images

Berat Mert Albaba, Sedat Ozer
Track 3: Computer Vision Robotics and Intelligent Systems
Fri 15 Jan 2021 at 16:00 in session PS T3.11

Auto-TLDR; SyNet: Combining Multi-Stage and Single-Stage Object Detection for Aerial Images

Recent advances in camera equipped drone applications and their widespread use increased the demand on vision based object detection algorithms for aerial images. Object detection process is inherently a challenging task as a generic computer vision problem, however, since the use of object detection algorithms on UAVs (or on drones) is relatively a new area, it remains as a more challenging problem to detect objects in aerial images. There are several reasons for that including: (i) the lack of large drone datasets including large object variance, (ii) the large orientation and scale variance in drone images when compared to the ground images, and (iii) the difference in texture and shape features between the ground and the aerial images. Deep learning based object detection algorithms can be classified under two main categories: (a) single-stage detectors and (b) multi-stage detectors. Both single-stage and multi-stage solutions have their advantages and disadvantages over each other. However, a technique to combine the good sides of each of those solutions could yield even a stronger solution than each of those solutions individually. In this paper, we propose an ensemble network, SyNet, that combines a multi-stage method with a single-stage one with the motivation of decreasing the high false negative rate of multi-stage detectors and increasing the quality of the single-stage detector proposals. As building blocks, CenterNet and Cascade R-CNN with pretrained feature extractors are utilized along with an ensembling strategy. We report the state of the art results obtained by our proposed solution on two different datasets: namely MS-COCO and visDrone with \%52.1 $mAP_{IoU = 0.75}$ is obtained on MS-COCO $val2017$ dataset and \%26.2 $mAP_{IoU = 0.75}$ is obtained on VisDrone $test-set$. Our code is available at: https://github.com/mertalbaba/SyNet}{https://github.com/mer talbaba/SyNet

#### Object Detection Model Based on Scene-Level Region Proposal Self-Attention

Yu Quan, Zhixin Li, Canlong Zhang, Huifang Ma
Track 3: Computer Vision Robotics and Intelligent Systems
Wed 13 Jan 2021 at 12:00 in session PS T3.4

Auto-TLDR; Exploiting Semantic Informations for Object Detection

The improvement of object detection performance is mostly focused on the extraction of local information near the region of interest in the image, which results in detection performance in this area being unable to achieve the desired effect. First, a depth-wise separable convolution network(D_SCNet-127 R-CNN) is built on the backbone network. Considering the importance of scene and semantic informations for visual recognition, the feature map is sent into the branch of the semantic segmentation module, region proposal network module, and the region proposal self-attention module to build the network of scene-level and region proposal self-attention module. Second, a deep reinforcement learning was utilized to achieve accurate positioning of border regression, and the calculation speed of the whole model was improved through implementing a light-weight head network. This model can effectively solve the limitation of feature extraction in traditional object detection and obtain more comprehensive detailed features. The experimental verification on MSCOCO17, VOC12, and Cityscapes datasets shows that the proposed method has good validity and scalability.

#### P2 Net: Augmented Parallel-Pyramid Net for Attention Guided Pose Estimation

Luanxuan Hou, Jie Cao, Yuan Zhao, Haifeng Shen, Jian Tang, Ran He
Track 2: Biometrics, Human Analysis and Behavior Understanding
Fri 15 Jan 2021 at 15:00 in session PS T2.5

Auto-TLDR; Parallel-Pyramid Net with Partial Attention for Human Pose Estimation

The target of human pose estimation is to determine the body parts and joint locations of persons in the image. Angular changes, motion blur and occlusion etc. in the natural scenes make this task challenging, while some joints are more difficult to be detected than others. In this paper, we propose an augmented Parallel-Pyramid Net (P^2Net) with an partial attention module. During data preprocessing, we proposed a differentiable auto data augmentation (DA^2) method in which sequences of data augmentations are formulated as a trainable and operational Convolution Neural Network (CNN) component. DA^2 improves the training efficiency and effectiveness. A parallel pyramid structure is followed to compensate the information loss introduced by the network. For the information loss problem in the backbone network, we optimize the backbone network by adopting a new parallel structure without increasing the overall computational complexity. To further refine the predictions after completion of global predictions, an Partial Attention Module (PAM) is defined to extract weighted features from different scale feature maps generated by the parallel pyramid structure. Compared with the traditional up-sampling refining, PAM can better capture the relationship between channels. Experiments corroborate the effectiveness of our proposed method. Notably, our method achieves the best performance on the challenging MSCOCO and MPII datasets.

#### Detective: An Attentive Recurrent Model for Sparse Object Detection

Amine Kechaou, Manuel Martinez, Monica Haurilet, Rainer Stiefelhagen
Track 3: Computer Vision Robotics and Intelligent Systems
Fri 15 Jan 2021 at 16:00 in session PS T3.11

Auto-TLDR; Detective: An attentive object detector that identifies objects in images in a sequential manner

In this work, we present Detective – an attentive object detector that identifies objects in images in a sequential manner. Our network is based on an encoder-decoder architecture, where the encoder is a convolutional neural network, and the decoder is a convolutional recurrent neural network coupled with an attention mechanism. At each iteration, our decoder focuses on the relevant parts of the image using an attention mechanism, and then estimates the object’s class and the bounding box coordinates. Current object detection models generate dense predictions and rely on post-processing to remove duplicate predictions. Detective is a sparse object detector that generates a single bounding box per object instance. However, training a sparse object detector is challenging, as it requires the model to reason at the instance level and not just at the class and spatial levels. We propose a training mechanism based on the Hungarian Algorithm and a loss that balances the localization and classification tasks. This allows Detective to achieve promising results on the PASCAL VOC object detection dataset. Our experiments demonstrate that sparse object detection is possible and has a great potential for future developments in applications where the order of the objects to be predicted is of interest.

#### VTT: Long-Term Visual Tracking with Transformers

Tianling Bian, Yang Hua, Tao Song, Zhengui Xue, Ruhui Ma, Neil Robertson, Haibing Guan
Track 3: Computer Vision Robotics and Intelligent Systems
Wed 13 Jan 2021 at 14:00 in session OS T3.2

Auto-TLDR; Visual Tracking Transformer with transformers for long-term visual tracking

Long-term visual tracking is a challenging problem. State-of-the-art long-term trackers, e.g., GlobalTrack, utilize region proposal networks (RPNs) to generate target proposals. However, the performance of the trackers is affected by occlusions and large scale or ratio variations. To address these issues, in this paper, we are the first to propose a novel architecture with transformers for long-term visual tracking. Specifically, the proposed Visual Tracking Transformer (VTT) utilizes a transformer encoder-decoder architecture for aggregating global information to deal with occlusion and large scale or ratio variation. Furthermore, it also shows better discriminative power against instance-level distractors without the need for extra labeling and hard-sample mining. We conduct extensive experiments on three largest long-term tracking dataset and have achieved state-of-the-art performance.

#### Attention Pyramid Module for Scene Recognition

Zhinan Qiao, Xiaohui Yuan, Chengyuan Zhuang, Abolfazl Meyarian
Track 3: Computer Vision Robotics and Intelligent Systems
Tue 12 Jan 2021 at 17:00 in session PS T3.3

Auto-TLDR; Attention Pyramid Module for Multi-Scale Scene Recognition

The unrestricted open vocabulary and diverse substances of scenery images bring significant challenges to scene recognition. However, most deep learning architectures and attention methods are developed on general-purpose datasets and omit the characteristics of scene data. In this paper, we exploit the attention pyramid module (APM) to tackle the predicament of scene recognition. Our method streamlines the multi-scale scene recognition pipeline, learns comprehensive scene features at various scales and locations, addresses the interdependency among scales, and further assists feature re-calibration as well as aggregation process. APM is extremely light-weighted and can be easily plugged into existing network architectures in a parameter-efficient manner. By simply integrating APM into ResNet-50, we obtain a 3.54\% boost in terms of top-1 accuracy on the benchmark scene dataset. Comprehensive experiments show that APM achieves better performance comparing with state-of-the-art attention methods using significant less computation budget. Code and pre-trained models will be made publicly available.

#### ACRM: Attention Cascade R-CNN with Mix-NMS for Metallic Surface Defect Detection

Junting Fang, Xiaoyang Tan, Yuhui Wang
Track 3: Computer Vision Robotics and Intelligent Systems
Thu 14 Jan 2021 at 14:00 in session PS T3.8

Auto-TLDR; Attention Cascade R-CNN with Mix Non-Maximum Suppression for Robust Metal Defect Detection

Metallic surface defect detection is of great significance in quality control for production. However, this task is very challenging due to the noise disturbance, large appearance variation, and the ambiguous definition of the defect individual. Traditional image processing methods are unable to detect the damaged region effectively and efficiently. In this paper, we propose a new defect detection method, Attention Cascade R-CNN with Mix-NMS (ACRM), to classify and locate defects robustly. Three submodules are developed to achieve this goal: 1) a lightweight attention block is introduced, which can improve the ability in capture global and local feature both in the spatial and channel dimension; 2) we firstly apply the cascade R-CNN to our task, which exploits multiple detectors to sequentially refine the detection result robustly; 3) we introduce a new method named Mix Non-Maximum Suppression (Mix-NMS), which can significantly improve its ability in filtering the redundant detection result in our task. Extensive experiments on a real industrial dataset show that ACRM achieves state-of-the-art results compared to the existing methods, demonstrating the effectiveness and robustness of our detection method.

#### VGG-Embedded Adaptive Layer-Normalized Crowd Counting Net with Scale-Shuffling Modules

Dewen Guo, Jie Feng, Bingfeng Zhou
Track 1: Artificial Intelligence, Machine Learning for Pattern Analysis
Tue 12 Jan 2021 at 15:00 in session PS T1.1

Crowd counting is widely used in real-time congestion monitoring and public security. Due to the limited data, many methods have little ability to be generalized because the differences between feature domains are not taken into consideration. We propose VGG-embedded adaptive layer normalization (VadaLN) to filter the features that irrelevant to the counting tasks in order that the counting results should not be affected by the image quality, color or illumination. VadaLN is implemented on the pretrained VGG-16 backbone. There is no additional learning parameters required through our method. VadaLN incoporates the proposed scale-shuffling modules (SSM) to relax the distortions in upsampling operations. Besides, non-aligned training methdology for the estimation of density maps is leveraged by an adversarial contextual loss (ACL) to improve the counting performance. Based on the proposed method, we construct an end-to-end trainable baseline model without bells and whistles, namely VadaLNet, which outperforms several recent state-of-the-art methods on commonly used challenging standard benchmarks. The intermediate scale-shuffled results are combined to formulate a scale-complementary strategy as a more powerful network, namely as VadaLNeSt. We implement VadaLNeSt on standard benchmarks, e.g. ShanghaiTech (Part A & Part B), UCF_CC_50, and UCF_QNRF, to show the superiority of our method.

#### Enhanced Feature Pyramid Network for Semantic Segmentation

Mucong Ye, Ouyang Jinpeng, Ge Chen, Jing Zhang, Xiaogang Yu
Track 1: Artificial Intelligence, Machine Learning for Pattern Analysis
Fri 15 Jan 2021 at 15:00 in session PS T1.14

Auto-TLDR; EFPN: Enhanced Feature Pyramid Network for Semantic Segmentation

Multi-scale feature fusion has been an effective way for improving the performance of semantic segmentation. However, current methods generally fail to consider the semantic gaps between the shallow (low-level) and deep (high-level) features and thus the fusion methods may not be optimal. In this paper, to address the issues of the semantic gap between the feature from different layers, we propose a unified framework based on the U-shape encoder-decoder architecture, named Enhanced Feature Pyramid Network (EFPN). Specifically, the semantic enhancement module (SEM), boundary extraction module (BEM), and context aggregation model (CAM) are incorporated into the decoder network to improve the robustness of the multi-level features aggregation. In addition, a global fusion model (GFM) in encoder branch is proposed to capture more semantic information in the deep layers and effectively transmit the high-level semantic features to each layer. Extensive experiments are conducted and the results show that the proposed framework achieves the state-of-the-art results on three public datasets, namely PASCAL VOC 2012, Cityscapes, and PASCAL Context. Furthermore, we also demonstrate that the proposed method is effective for other visual tasks that require frequent fusing features and upsampling.

#### Dynamic Low-Light Image Enhancement for Object Detection Via End-To-End Training

Haifeng Guo, Yirui Wu, Tong Lu
Track 5: Image and Signal Processing
Thu 14 Jan 2021 at 12:00 in session PS T5.5

Auto-TLDR; Object Detection using Low-Light Image Enhancement for End-to-End Training

Object detection based on convolutional neural networks is a hot research topic in computer vision. The illumination component in the image has a great impact on object detection, and it will cause a sharp decline in detection performance under low-light conditions. Using low-light image enhancement technique as a pre-processing mechanism can improve image quality and obtain better detection results.However, due to the complexity of low-light environments, the existing enhancement methods may have negative effects on some samples. Therefore, it is difficult to improve the overall detection performance in low-light conditions. In this paper, our goal is to use image enhancement to improve object detection performance rather than perceptual quality for humans. We propose a novel framework that combines low-light enhancement and object detection for end-to-end training. The framework can dynamically select different enhancement subnetworks for each sample to improve the performance of the detector. Our proposed method consists of two stage: the enhancement stage and the detection stage. The enhancement stage dynamically enhances the low-light images under the supervision of several enhancement methods and output corresponding weights. During the detection stage, the weights offers information on object classification to generate high-quality region proposals and in turn result in accurate detection. Our experiments present promising results, which show that the proposed method can significantly improve the detection performance in low-light environment.

#### A Novel Region of Interest Extraction Layer for Instance Segmentation

Leonardo Rossi, Akbar Karimi, Andrea Prati
Track 1: Artificial Intelligence, Machine Learning for Pattern Analysis
Wed 13 Jan 2021 at 16:30 in session PS T1.8

Auto-TLDR; Generic RoI Extractor for Two-Stage Neural Network for Instance Segmentation

Given the wide diffusion of deep neural network architectures for computer vision tasks, several new applications are nowadays more and more feasible. Among them, a particular attention has been recently given to instance segmentation, by exploiting the results achievable by two-stage networks (such as Mask R-CNN or Faster R-CNN), derived from R-CNN. In these complex architectures, a crucial role is played by the Region of Interest (RoI) extraction layer, devoted to extract a coherent subset of features from a single Feature Pyramid Network (FPN) layer attached on top of a backbone. This paper is motivated by the need to overcome to the limitations of existing RoI extractors which select only one (the best) layer from FPN. Our intuition is that all the layers of FPN retain useful information. Therefore, the proposed layer (called Generic RoI Extractor - GRoIE) introduces non-local building blocks and attention mechanisms to boost the performance. A comprehensive ablation study at component level is conducted to find the best set of algorithms and parameters for the GRoIE layer. Moreover, GRoIE can be integrated seamlessly with every two-stage architecture for both object detection and instance segmentation tasks. Therefore, the improvements brought by the use of GRoIE in different state-of-the-art architectures are also evaluated. The proposed layer leads up to gain a 1.1% AP on bounding box detection and 1.7% AP on instance segmentation. The code is publicly available on GitHub repository at https://github.com/IMPLabUniPr/mmdetection-groie

#### Yolo+FPN: 2D and 3D Fused Object Detection with an RGB-D Camera

Ya Wang
Track 3: Computer Vision Robotics and Intelligent Systems
Tue 12 Jan 2021 at 17:00 in session PS T3.3

Auto-TLDR; Yolo+FPN: Combining 2D and 3D Object Detection for Real-Time Object Detection

In this paper we propose a new deep neural network system, called Yolo+FPN, which fuses both 2D and 3D object detection algorithms to achieve better real-time object detection results and faster inference speed, to be used on real robots. Finding an optimized fusion strategy to efficiently combine 3D object detection with 2D detection information is useful and challenging for both indoor and outdoor robots. In order to satisfy real-time requirements, a trade-off between accuracy and efficiency is needed. We not only have improved training and test accuracies and lower mean losses on the KITTI object detection benchmark, but also achieve better average precision on 3D detection of all classes in three levels of difficulty. Also, we implemented Yolo+FPN system using an RGB-D camera, and compared the speed of 2D and 3D object detection using different GPUs. For the real implementation of both indoor and outdoor scenes, we focus on person detection, which is the most challenging and important among the three classes.

#### Siamese Dynamic Mask Estimation Network for Fast Video Object Segmentation

Dexiang Hong, Guorong Li, Kai Xu, Li Su, Qingming Huang
Track 3: Computer Vision Robotics and Intelligent Systems
Thu 14 Jan 2021 at 16:00 in session PS T3.9

Auto-TLDR; Siamese Dynamic Mask Estimation for Video Object Segmentation

Video object segmentation(VOS) has been a fundamental topic in recent years, and many deep learning-based methods have achieved state-of-the-art performance on multiple benchmarks. However, most of these methods rely on pixel-level matching between the template and the searched frames on the whole image while the targets only occupy a small region. Calculating on the entire image brings lots of additional computation cost. Besides, the whole image may contain some distracting information resulting in many false-positive matching points. To address this issue, motivated by one-stage instance object segmentation methods, we propose an efficient siamese dynamic mask estimation network for fast video object segmentation. The VOS is decoupled into two tasks, i.e. mask feature learning and dynamic kernel prediction. The former is responsible for learning high-quality features to preserve structural geometric information, and the latter learns a dynamic kernel which is used to convolve with the mask feature to generate a mask output. We use Siamese neural network as a feature extractor and directly predict masks after correlation. In this way, we can avoid using pixel-level matching, making our framework more simple and efficient. Experiment results on DAVIS 2016 /2017 datasets show that our proposed methods can run at 35 frames per second on NVIDIA RTX TITAN while preserving competitive accuracy.

#### You Ought to Look Around: Precise, Large Span Action Detection

Ge Pan, Zhang Han, Fan Yu, Yonghong Song, Yuanlin Zhang, Han Yuan
Track 3: Computer Vision Robotics and Intelligent Systems
Thu 14 Jan 2021 at 16:00 in session OS T3.4

Auto-TLDR; YOLA: Local Feature Extraction for Action Localization with Variable receptive field

For the action localization task, pre-defined action anchors are the cornerstone of mainstream techniques. State-of-the-art models mostly rely on a dense segmenting scheme, where anchors are sampled uniformly over the temporal domain with a predefined set of scales. However, it is not sufficient because action duration varies greatly. Therefore, it is necessary for the anchors or proposals to have a variable receptive field. In this paper, we propose a method called YOLA (You Ought to Look Around) which includes three parts: 1) a robust backbone SPN-I3D for extracting spatio-temporal features. In this part, we employ a stronger backbone I3D with SPN (Segment Pyramid Network) instead of C3D to obtain multi-scale features; 2) a simple but useful feature fusion module named LFE (Local Feature Extraction). Compared with the fully connected layer and global average pooling, our LFE model is more advantageous for network to fit and fuse features. 3) a new feature segment aligning method called TPGC (Two Pathway Graph Convolution), which allows one proposal to leverage semantic features of adjacent proposals to update its content and make sure the proposals have a variable receptive field. YOLA add only a small overhead to the baseline network, and is easy to train in an end-to-end manner, running at a speed of 1097 fps. YOLA achieves a mAP of 58.3%, outperforming all existing models including both RGB-based and two stream on THUMOS'14, and achieves competitive results on ActivityNet 1.3.

#### GSTO: Gated Scale-Transfer Operation for Multi-Scale Feature Learning in Semantic Segmentation

Zhuoying Wang, Yongtao Wang, Zhi Tang, Yangyan Li, Ying Chen, Haibin Ling, Weisi Lin
Track 3: Computer Vision Robotics and Intelligent Systems
Thu 14 Jan 2021 at 14:00 in session PS T3.8

Auto-TLDR; Gated Scale-Transfer Operation for Semantic Segmentation

Existing CNN-based methods for semantic segmentation heavily depend on multi-scale features to meet the requirements of both semantic comprehension and detail preservation. State-of-the-art segmentation networks widely exploit conventional scale-transfer operations, i.e., up-sampling and down-sampling to learn multi-scale features. In this work, we find that these operations lead to scale-confused features and suboptimal performance because they are spatial-invariant and directly transit all feature information cross scales without spatial selection. To address this issue, we propose the Gated Scale-Transfer Operation (GSTO) to properly transit spatial-filtered features to another scale. Specifically, GSTO can work either with or without extra supervision. Unsupervised GSTO is learned from the feature itself while the supervised one is guided by the supervised probability matrix. Both forms of GSTO are lightweight and plug-and-play, which can be flexibly integrated into networks or modules for learning better multi-scale features. In particular, by plugging GSTO into HRNet, we get a more powerful backbone (namely GSTO-HRNet) for pixel labeling, and it achieves new state-of-the-art results on multiple benchmarks for semantic segmentation including Cityscapes, LIP and Pascal Context, with negligible extra computational cost. Moreover, experiment results demonstrate that GSTO can also significantly boost the performance of multi-scale feature aggregation modules like PPM and ASPP.

#### HANet: Hybrid Attention-Aware Network for Crowd Counting

Xinxing Su, Yuchen Yuan, Xiangbo Su, Zhikang Zou, Shilei Wen, Pan Zhou
Track 1: Artificial Intelligence, Machine Learning for Pattern Analysis
Wed 13 Jan 2021 at 14:00 in session OS T1.4

Auto-TLDR; HANet: Hybrid Attention-Aware Network for Crowd Counting with Adaptive Compensation Loss

An essential yet challenging issue in crowd counting is the diverse background variations under complicated real-life environments, which makes attention based methods favorable in recent years. However, most existing methods only rely on first-order attention schemes (e.g. 2D position-wise attention), while ignoring the higher-order information within the congested scenes completely. In this paper, we propose a hybrid attention-aware network (HANet) with a high-order attention module (HAM) and an adaptive compensation loss (ACLoss) to tackle this problem. On the one hand, the HAM applies 3D attention to capture the subtle discriminative features around each people in the crowd. On the other hand, with the distributed supervision, the ACLoss exploits the prior knowledge from higher-level stages to guide the density map prediction at a lower level. The proposed HANet is then established with HAM and ACLoss working as different roles and promoting each other. Extensive experimental results show the superiority of our HANet against the state-of-the-arts on three challenging benchmarks.

#### Hierarchical Head Design for Object Detectors

Shivang Agarwal, Frederic Jurie
Track 3: Computer Vision Robotics and Intelligent Systems
Wed 13 Jan 2021 at 16:30 in session PS T3.5

Auto-TLDR; Hierarchical Anchor for SSD Detector

The notion of anchor plays a major role in modern detection algorithms such as the Faster-RCNN or the SSD detector. Anchors relate the features of the last layers of the detector with bounding boxes containing objects in images. Despite their importance, the literature on object detection has not paid real attention to them. The motivation of this paper comes from the observations that (i) each anchor learns to classify and regress candidate objects independently (ii) insufficient examples are available for each anchor in case of small-scale datasets. This paper addresses these questions by proposing a novel hierarchical head for the SSD detector. The new design has the added advantage of no extra weights, as compared to the original design at inference time, while improving detectors performance for small size training sets. Improved performance on PASCAL-VOC and state-of-the-art performance on FlickrLogos-47 validate the method. We also show when the proposed design does not give additional performance gain over the original design.

#### TCATD: Text Contour Attention for Scene Text Detection

Ziling Hu, Wu Xingjiao, Jing Yang
Track 4: Document and Media Analysis
Thu 14 Jan 2021 at 12:00 in session PS T4.3

Auto-TLDR; Text Contour Attention Text Detector

Segmentation-based approaches have enabled state-of-the-art performance in long or curved text detection tasks. However, false detection still is a challenge when two text instances are close to each other. To address this problem, in this paper, we propose a Text Contour Attention Text Detector (TCATD), which can locate scene text with arbitrary orientation and shape accurately. Different from previous work, TCATD focus on text contour map (TC), text center intensity map (TCI) and text kernel maps (TK). The TC can introduce text contour information, the TCI can help to learn the accurate text segmentation and the TK can generate the complete shape of text instances. Besides, we propose a Text Contour Attention Module to deal with contour information. After the Text Contour Attention Module, TC, TCI and TK will be obtained. Extensive experiments on ICDAR2015, CTW1500 and Total-Text demonstrate that the proposed method achieves the state-of-the-art performance.

#### Deep Real-Time Hand Detection Using CFPN on Embedded Systems

Pirdiansyah Hendri, Jun-Wei Hsieh, Ping Yang Chen
Track 2: Biometrics, Human Analysis and Behavior Understanding
Thu 14 Jan 2021 at 12:00 in session PS T2.4

Auto-TLDR; Concatenated Feature Pyramid Network for Small Hand Detection on Embedded Devices

Real-time HI (Human Interface) systems need accurate and efficient hand detection models to meet the limited resources in budget, dimension, memory, computing, and electric power. In recent years, object detection became a less challenging task with the latest deep CNN-based state-of-the-art models, i.e., RCNN, SSD, and YOLO; however, these models cannot provide the desired efficiency and accuracy for HI systems on embedded devices due to their complex time-consuming architecture. In addition, the detection of small hands (<30x30 pixels) is still a challenging task for all the above existing methods. Thus, we propose a shallow model named Concatenated Feature Pyramid Network (CFPN) to provide above mentioned performance for small hand detection. The superiority of CFPN is confirmed on a HandFlow dataset with mAP:0.5 of 95.6 and FPS of 33 on Nvidia TX2. The COCO dataset is also used to compare with other state-of-the-art method and shows the highest efficiency and accuracy with the proposed CFPN model. Thus we conclude that the proposed model is useful for real-life small hand detection on embedded devices.

#### CDeC-Net: Composite Deformable Cascade Network for Table Detection in Document Images

Madhav Agarwal, Ajoy Mondal, C. V. Jawahar
Track 4: Document and Media Analysis
Tue 12 Jan 2021 at 14:00 in session OS T4.1

Auto-TLDR; CDeC-Net: An End-to-End Trainable Deep Network for Detecting Tables in Document Images

Localizing page elements/objects such as tables, figures, equations, etc. is the primary step in extracting information from document images. We propose a novel end-to-end trainable deep network, (CDeC-Net) for detecting tables present in the documents. The proposed network consists of a multistage extension of Mask R-CNN with a dual backbone having deformable convolution for detecting tables varying in scale with high detection accuracy at higher IoU threshold. We empirically evaluate CDeC-Net on all the publicly available benchmark datasets— ICDAR-2013, ICDAR-2017, ICDAR-2019, UNLV, Marmot, PubLayNet, TableBank, and IIIT-AR-13K —with extensive experiments. Our solution has three important properties:(i) a single trained model CDeC-Net‡ performs well across all the popular benchmark datasets; (ii) we report excellent performances across multiple, including higher, thresholds of IoU; (iii) by following the same protocol of the recent papers for each of the benchmarks, we consistently demonstrate the superior quantitative performance. Our code and models will be publicly released for enabling reproducibility of the results.

#### Feature Pyramid Hierarchies for Multi-Scale Temporal Action Detection

Jiayu He, Guohui Li, Jun Lei
Track 1: Artificial Intelligence, Machine Learning for Pattern Analysis
Thu 14 Jan 2021 at 12:00 in session PS T1.10

Auto-TLDR; Temporal Action Detection using Pyramid Hierarchies and Multi-scale Feature Maps

Temporal action detection is a challenging but promising task in video content analysis. It is in great demand in the field of public safety. The main difficulty of the task is precisely localizing activities in the video especially those short duration activities. And most of the existing methods can not achieve a satisfactory detection result. Our method addresses a key point to improve detection accuracy, which is to use multi-scale feature maps for regression and classification. In this paper, we introduce a novel network based on classification following proposal framework. In our network, a 3D feature pyramid hierarchies is built to enhance the ability of detecting short duration activities. The input RGB/Flow frames are first encoded by a 3D feature pyramid hierarchies, and this subnet produces multi-level feature maps. Then temporal proposal subnet uses these features to pick out proposals which might contain activity segments. Finally a pyramid region of interest (RoI) pooling pipeline and two fully connected layers reuse muti-level feature maps to refine the temporal boundaries of proposals and classify them. We use late feature fusion scheme to combine RGB and Flow information. The network is trained end-to-end and we evaluate it in THUMOS'14 dataset. Our network achieves a good result among typical methods. A further ablation test demonstrate that pyramid hierarchies is effective to improve detecting short duration activity segments.

#### Enhanced Vote Network for 3D Object Detection in Point Clouds

Min Zhong, Gang Zeng
Track 3: Computer Vision Robotics and Intelligent Systems
Thu 14 Jan 2021 at 12:00 in session PS T3.7

Auto-TLDR; A Vote Feature Enhancement Network for 3D Bounding Box Prediction

In this work, we aim to estimate 3D bounding boxes by voting to object centers and then groups and aggregates the votes to generate 3D box proposals and semantic classes of objects. However, due to the sparse and unstructured nature of the point clouds, we face some challenges when directly predicting bounding box from the vote feature: the sparse vote feature may lack some necessary semantic and context information. To address the challenges, we propose a vote feature enhancement network that aims to encode semantic-aware information and aggravate global context for the vote feature. Specifically, we learn the point-wise semantic information and supplement it to the vote feature, and we also encode the pairwise relations to collect the global context. Experiments on two large datasets of real 3D scans, ScanNet and SUN RGB-D, demonstrate that our method can achieve excellent 3D detection results.

#### Real-Time Semantic Segmentation Via Region and Pixel Context Network

Yajun Li, Yazhou Liu, Quansen Sun
Track 3: Computer Vision Robotics and Intelligent Systems
Thu 14 Jan 2021 at 12:00 in session PS T3.7

Auto-TLDR; A Dual Context Network for Real-Time Semantic Segmentation

Real-time semantic segmentation is a challenging task as both segmentation accuracy and inference speed need to be considered at the same time. In this paper, we present a Dual Context Network (DCNet) to address this challenge. It contains two independent sub-networks: Region Context Network and Pixel Context Network. Region Context Network is main network with low-resolution input and feature re-weighting module to achieve sufficient receptive field. Meanwhile, Pixel Context Network with location attention module to capture the location dependencies of each pixel for assisting the main network to recover spatial detail. A contextual feature fusion is introduced to combine output features of these two sub-networks. The experiments show that DCNet can achieve high-quality segmentation while keeping a high speed. Specifically, for Cityscapes test dataset, we achieve 76.1% Mean IOU with the speed of 82 FPS on a single GTX 2080Ti GPU when using ResNet50 as backbone, and 71.2% Mean IOU with the speed of 142 FPS when using ResNet18 as backbone.

#### One-Stage Multi-Task Detector for 3D Cardiac MR Imaging

Weizeng Lu, Xi Jia, Wei Chen, Nicolò Savioli, Antonio De Marvao, Linlin Shen, Declan O'Regan, Jinming Duan
Track 5: Image and Signal Processing
Thu 14 Jan 2021 at 12:00 in session PS T5.5

Auto-TLDR; Multi-task Learning for Real-Time, simultaneous landmark location and bounding box detection in 3D space

Fast and accurate landmark location and bounding box detection are important steps in 3D medical imaging. In this paper, we propose a novel multi-task learning framework, for real-time, simultaneous landmark location and bounding box detection in 3D space. Our method extends the famous single-shot multibox detector (SSD) from single-task learning to multi-task learning and from 2D to 3D. Furthermore, we propose a post-processing approach to refine the network landmark output, by averaging the candidate landmarks. Owing to these settings, the proposed framework is fast and accurate. For 3D cardiac magnetic resonance (MR) images with size 224 × 224 × 64, our framework runs about 128 volumes per second (VPS) on GPU and achieves 6.75mm average point-to-point distance error for landmark location, which outperforms both state-of-the-art and baseline methods. We also show that segmenting the 3D image cropped with the bounding box results in both improved performance and efficiency.