In the last few years, several techniques for facial manipulation in videos have been successfully developed and made available to the masses (i.e., FaceSwap, deepfake, etc.). These methods enable anyone to easily edit faces in video sequences with incredibly realistic results and a very little effort. Despite the usefulness of these tools in many fields, if used maliciously, they can have a significantly bad impact on society (e.g., fake news spreading, cyber bullying through fake revenge porn). The ability of objectively detecting whether a face has been manipulated in a video sequence is then a task of utmost importance. In this paper, we tackle the problem of face manipulation detection in video sequences targeting modern facial manipulation techniques. In particular, we study the ensembling of different trained Convolutional Neural Network (CNN) models. In the proposed solution, different models are obtained starting from a base network (i.e., EfficientNetB4) making use of two different concepts: (i) attention layers; (ii) siamese training. We show that combining these networks leads to promising face manipulation detection results on two publicly available datasets with more than 119000 videos.

Visual dialog is a typical multi-modal task which involves both vision and language. Nowadays, it faces two major difficulties. In this paper, we propose Integrating Historical States and Co-attention (HSCA) for visual dialog to solve them. It includes two main modules, Co-ATT and MATCH. Specifically, the main purpose of the Co-ATT module is to guide the image with questions and answers in the early stage to get more specific objects. It tackles the temporal sequence issue in historical information which may influence the precise answer for multi-round questions. The MATCH module is, based on a question with pronouns, to retrieve the best matching historical information block. It overcomes the visual reference problem which requires to solve pronouns referring to unknowns in the text message and then to locate the objects in the given image. We quantitatively and qualitatively evaluate our model on VisDial v1.0, at the same time, ablation studies are carried out. The experimental results demonstrate that HSCA outperforms the state-of-the-art methods in many aspects.

Deep feedforward convolutional neural networks (CNNs) perform well in the saliency prediction of omnidirectional images (ODIs), and have become the leading class of candidate models of the visual processing mechanism in the primate ventral stream. These CNNs have evolved from shallow network architecture to extremely deep and branching architecture to achieve superb performance in various vision tasks, yet it is unclear how brain-like they are. In particular, these deep feedforward CNNs are difficult to mapping to ventral stream structure of the brain visual system due to their vast number of layers and missing biologically-important connections, such as recurrence. To tackle this issue, some brain-like shallow neural networks are introduced. In this paper, we propose a novel brain-like network model for saliency prediction of head fixations on ODIs. Specifically, our proposed model consists of three modules: a CORnet-S module, a template feature extraction module and a ranking attention module (RAM). The CORnet-S module is a lightweight artificial neural network (ANN) with four anatomically mapped areas (V1, V2, V4 and IT) and it can simulate the visual processing mechanism of ventral visual stream in the human brain. The template features extraction module is introduced to extract attention maps of ODIs and provide guidance for the feature ranking in the following RAM module. The RAM module is used to rank and select features that are important for fine-grained saliency prediction. Extensive experiments have validated the effectiveness of the proposed model in predicting saliency maps of ODIs, and the proposed model outperforms other state-of-the-art methods with similar scale.

In this paper, we propose a multi-attribute regression network (MARN) to investigate the problem of face reconstruction, especially in challenging cases when faces undergo large variations including severe poses, extreme expressions, and partial occlusions in unconstrained environments. The traditional 3DMM parametric regression method is absent from the learning of identity, expression, and attitude attributes, resulting in lacking geometric details in the reconstructed face. Our MARN method is to enable the network to better extract the feature information of face identity, expression, and pose attributes. We introduced identity, expression, and pose attribute loss functions to enhance the learning of details in each attribute. At the same time, we carefully design the geometric contour constraint loss function and use the constraints of sparse 2D face landmarks to improve the reconstructed geometric contour information. The experimental results show that our face reconstruction method has achieved significant results on the AFLW2000-3D and AFLW datasets compared with the most advanced methods. In addition, there has been a great improvement in dense face alignment. .

The current deep learning based visual tracking approaches have been very successful by learning the target classification and/or estimation model from a large amount of supervised training data in offline mode. However, most of them can still fail in tracking objects due to some more challenging issues such as dense distractor objects, confusing background, motion blurs, and so on. Inspired by the human ``visual tracking'' capability which leverages motion cues to distinguish the target from the background, we propose a Two-Stream Residual Convolutional Network (TS-RCN) for visual tracking, which successfully exploits both appearance and motion features for model update. Our TS-RCN can be integrated with existing deep learning based visual trackers. To further improve the tracking performance, we adopt a ``wider'' residual network ResNeXt as its feature extraction backbone. To the best of our knowledge, TS-RCN is the first end-to-end trainable two-stream visual tracking system, which makes full use of both appearance and motion features of the target. We have extensively evaluated the TS-RCN on most widely used benchmark datasets including VOT2018, VOT2019, and GOT-10K. The experiment results have successfully demonstrated that our two-stream model can greatly outperform the appearance based tracker, and it also achieves state-of-the-art performance. The tracking system can run at up to 38.1 FPS.

With the prosperous development of Convolutional Neural Networks, currently they can perform excellently on visual understanding tasks when the input images are high quality and common quality images. However, large degradation in performance always occur when the input images are low quality images. In this paper, we propose a new super-resolution method in order to improve the classification performance for low-resolution images. In an image, the regions in which pixel values vary dramatically contain more abundant high frequency contents compared to other parts. Based on this fact, we design a weight map and integrate it with a super-resolution CNN training framework. During the process of training, this weight map can find out positions of the high frequency pixels in ground truth high-resolution images. After that, the pixel-level loss function takes effect only at these found positions to minimize the difference between reconstructed high-resolution images and ground truth high-resolution images. Compared with other state-of-the-art super-resolution methods, the experiment results show that our method can recover more high-frequency contents in high-resolution image reconstructing, and better improve the classification accuracy after low-resolution image preprocessing.

Convolutional Neural Networks (CNNs) model long-range dependencies by deeply stacking convolution operations with small window sizes, which makes the optimizations difficult. This paper presents region-based non-local operation (RNL), a family of self-attention mechanisms, which can directly capture long-range dependencies without a deep stack of local operations. Given an intermediate feature map, our method recalibrates the feature at a position by aggregating information from the neighboring regions of all positions. By combining a channel attention module with the proposed RNL, we design an attention chain, which can be integrated into off-the-shelf CNNs for end-to-end training. We evaluate our method on two video classification benchmarks. The experimental result of our method outperforms other attention mechanisms, and we achieve state-of-the-art performance on Something-Something V1.

Traditional Chinese paintings pay more attention to ’Gongbi’ and ’Xieyi’ in artworks, which raises a challenging task to generate Chinese paintings from photos. ’Xieyi’ creates high-level conception for paintings, while ’Gongbi’ refers to portraying local details in paintings. This paper proposes an attentional wavelet network for photo to Chinese painting transferring. We first introduce wavelets to obtain high-level conception and local details in Chinese paintings via 2-D haar wavelet transform. Moreover, we design high-level transform stream and local enhancement stream to dispose high frequencies and low frequency respectively. Furthermore, we exploit self-attention mechanism to compatibly pick up high-level information which is used to remedy the missing details when reconstructing the Chinese painting. To advance our experiment, we set up a new dataset named P2ADataset, with diverse photos and Chinese paintings on famous mountains around China. Experimental results comparing with the state-of-the-art style transferring algorithms verify the effectiveness of the proposed method. We will release the codes and data to the public.

In this paper, a new 2D transform named Discrete Mirror Transform (DMT) is presented. The DMT is computed by decomposing a signal into its even and odd parts around an optimal location in a given direction so that the signal energy is maximally split between the two components. After minimizing the information required to regenerate the original signal by removing redundant structures, the process is iterated leading the signal energy to distribute into a continuously smaller set of coefficients. The DMT can be displayed as a binary tree, where each node represents the single (even or odd) signal derived from the decomposition in the previous level. An optimized version of the DMT (ODMT) is also introduced, by exploiting the possibility to choose different directions at which performing the decomposition. Experimental simulations have been carried out in order to test the sparsity properties of the DMT and ODMT when applied on images: referring to both transforms, the results show a superior performance with respect to the popular Discrete Cosine Transform (DCT) and Discrete Wavelet Transform (DWT) in terms of non-linear approximation.

We propose a novel network initialization method using Perlin noise for training image classification networks with a limited amount of data. Our main idea is to initialize the network parameters by solving an artificial noise classification problem, where the aim is to classify Perlin noise samples into their noise categories. Specifically, the proposed method consists of two steps. First, it generates Perlin noise samples with category labels defined based on noise complexity. Second, it solves a classification problem, in which network parameters are optimized to classify the generated noise samples. This method produces a reasonable set of initial weights (filters) for image classification. To the best of our knowledge, this is the first work to initialize networks by solving an artificial optimization problem without using any real-world images. Our experiments show that the proposed method outperforms conventional initialization methods on four image classification datasets.

Deep reinforcement learning (DRL) has reached super human levels in complexes tasks like game solving (Go, StarCraft II), and autonomous driving. However, it remains an open question whether DRL can reach human level in applications to financial problems and in particular in detecting pattern crisis and consequently dis-investing. In this paper, we present an innovative DRL framework consisting in two sub-networks fed respectively with portfolio strategies past performances and standard deviation as well as additional contextual features. The second sub network plays an important role as it captures dependencies with common financial indicators features like risk aversion, economic surprise index and correlations between assets that allows taking into account context based information. We compare different network architectures either using layers of convolutions to reduce network's complexity or LSTM block to capture time dependency and whether previous allocations is important in the modeling. We also use adversarial training to make the final model more robust. Results on test set show this approach substantially over-performs traditional portfolio optimization methods like Markovitz and is able to detect and anticipate crisis like the current Covid one.

Face recognition in unconstrained environments is challenging due to variations in illumination, quality of sensing, motion blur and etc. An individual’s face appearance can vary drastically under different conditions creating a gap between train (source) and varying test (target) data. The domain gap could cause decreased performance levels in direct knowledge transfer from source to target. Despite fine-tuning with domain specific data could be an effective solution, collecting and annotating data for all domains is extremely expensive. To this end, we propose a self-supervised domain learning (SSDL) scheme that trains on triplets mined from unlabelled data. A key factor in effective discriminative learning, is selecting informative triplets. Building on most confident predictions, we follow an “easy-to-hard” scheme of alternate triplet mining and self-learning. Comprehensive experiments on four different benchmarks show that SSDL generalizes well on different domains.

Accurately detecting craters in remotely sensed images is an important task when analysing the properties of planetary bodies. Commonly, only large craters in the range of several kilometres are detected. In this work we provide the first example of automatically detecting tiny craters in the range of several meters with the help of a deep neural network by using only a small set of annotated craters. Additionally, we propose a novel way to group overlapping detections and replace the commonly used non-maximum suppression with a probabilistic treatment. As a result, we receive valuable uncertainty estimates of the detections and the aggregated detections are shown to be vastly superior.

Recent machine learning methods use increasingly large deep neural networks to achieve state-of-the-art results in various tasks. Network quantization can effectively reduce computation and memory costs without modifying network structures, facilitating the deployment of deep neural networks (DNNs) on cloud and edge devices. However, most of the existing methods usually need time-consuming training or fine-tuning and access to the original training dataset that may be unavailable due to privacy or security concerns. In this paper, we present a novel method to achieve low-precision quantization of deep neural networks with limited data. Firstly, to reduce the complexity of per-channel quantization and the degeneration of per-layer quantization, we introduce group-wise quantization which separates the output channels into groups that each group is quantized separately. Secondly, to better distill knowledge from the pre-trained FP32 model with limited data, we introduce a two-stage knowledge distillation method that divides the optimization process into independent optimization stage and joint optimization stage to address the limitation of layer-wise supervision and global supervision. Extensive experiments on ImageNet2012 (ResNet18/50, ShuffleNetV2, and MobileNetV2) demonstrate that the proposed approach can significantly improve the quantization model's accuracy when only a few training samples are available. We further show that the method also extends to other computer vision architectures and tasks such as object detection and semantic segmentation.

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.

Manga character recognition is a key technology for manga character retrieval and verfication. This task is very challenging since the manga character images have a long-tailed distribution and large quality variations. Training models with cross-entropy softmax loss on such imbalanced data would introduce biases to feature and class weight norm. To handle this problem, we propose a novel dual loss which is the sum of two losses: dual ring loss and dual adaptive re-weighting loss. Dual ring loss combines weight and feature soft normalization and serves as a regularization term to softmax loss. Dual adaptive re-weighting loss re-weights softmax loss according to the norm of both feature and class weight. With the proposed losses, we have achieved encouraging results on Manga109 benchmark. Specifically, compared with the baseline softmax loss, our method improves the character retrieval mAP from 35.72% to 38.88% and the character verification accuracy from 87.00% to 88.50%.

The dimensionality reduction has been widely introduced to use the high-dimensional data for regression, classification, feature analysis, and visualization. As the one technique of dimensionality reduction, a stochastic neighbor embedding (SNE) was introduced. The SNE leads powerful results to visualize high-dimensional data by considering the similarity between the local Gaussian distributions of high and low-dimensional space. To improve the SNE, a t-distributed stochastic neighbor embedding (t-SNE) was also introduced. To visualize high-dimensional data, the t-SNE leads to more powerful and flexible visualization on 2 or 3-dimensional mapping than the SNE by using a t-distribution as the distribution of low-dimensional data. Recently, Uniform manifold approximation and projection (UMAP) is proposed as a dimensionality reduction technique. We present a novel technique called a q-Gaussian distributed stochastic neighbor embedding (q-SNE). The q-SNE leads to more powerful and flexible visualization on 2 or 3-dimensional mapping than the t-SNE and the SNE by using a q-Gaussian distribution as the distribution of low-dimensional data. The q-Gaussian distribution includes the Gaussian distribution and the t-distribution as the special cases with q=1.0 and q=2.0. Therefore, the q-SNE can also express the t-SNE and the SNE by changing the parameter q, and this makes it possible to find the best visualization by choosing the parameter q. We show the performance of q-SNE as visualization on 2-dimensional mapping and classification by k-Nearest Neighbors (k-NN) classifier in embedded space compared with SNE, t-SNE, and UMAP by using the datasets MNIST, COIL-20, OlivettiFaces, FashionMNIST, and Glove.

Most approaches to deep neural network compression via pruning either evaluate a filter’s importance using its weights or optimize an alternative objective function with sparsity constraints. While these methods offer a useful way to approximate contributions from similar filters, they often either ignore the dependency between layers or solve a more difficult optimization objective than standard cross-entropy. Our method, Mutual Information-based Neuron Trimming (MINT), approaches deep compression via pruning by enforcing sparsity based on the strength of the relationship between filters of adjacent layers, across every pair of layers. The relationship is calculated using conditional geometric mutual information which evaluates the amount of similar information exchanged between the filters using a graph-based criterion. When pruning a network, we ensure that retained filters contribute the majority of the information towards succeeding layers which ensures high performance. Our novel approach outperforms existing state-of-the-art compression-via-pruning methods on the standard benchmarks for this task: MNIST, CIFAR-10, and ILSVRC2012, across a variety of network architectures. In addition, we discuss our observations of a common denominator between our pruning methodology’s response to adversarial attacks and calibration statistics when compared to the original network.

Cross-modality RGB-infrared (RGB-IR) person re-identification (Re-ID) is a challenging research topic due to the heterogeneity of RGB and infrared images. In this paper, we aim to find some auxiliary modalities, which are homologous with the visible or infrared modalities, to help reduce the modality discrepancy caused by heterogeneous images. Accordingly, a new base-derivative framework is proposed, where base refers to the original visible and infrared modalities, and derivative refers to the two auxiliary modalities that are derived from base. In the proposed framework, the double-modality cross-modal learning problem is reformulated as a four-modality one. After that, the images of all the base and derivative modalities are fed into the feature learning network. With the doubled input images, the learned person features become more discriminative. Furthermore, the proposed framework is optimized by the enhanced intra- and cross-modality constraints with the assistance of two derivative modalities. Experimental results on two publicly available datasets SYSU-MM01 and RegDB show that the proposed method outperforms the other state-of-the-art methods. For instance, we achieve a gain of over 13\% in terms of both Rank-1 and mAP on RegDB dataset.

Few-shot segmentation presents a significant challenge for semantic scene understanding under limited supervision. Namely, this task targets at generalizing the segmentation ability of the model to new categories given a few samples. In order to obtain complete scene information, we extend the RGB-centric methods to take advantage of complementary depth information. In this paper, we propose a two-stream deep neural network based on metric learning. Our method, known as RDNet, learns class-specific prototype representations within RGB and depth embedding spaces, respectively. The learned prototypes provide effective semantic guidance on the corresponding RGB and depth query image, leading to more accurate performance. Moreover, we build a novel outdoor scene dataset, known as Cityscapes-3i, using labeled RGB images and depth images from the Cityscapes dataset. We also perform ablation studies to explore the effective use of depth information in few-shot segmentation tasks. Experiments on Cityscapes-3i show that our method achieves promising results with visual and complementary geometric cues from only a few labeled examples.