#### Attentive Visual Semantic Specialized Network for Video Captioning

Jesus Perez-Martin, Benjamin Bustos, Jorge Pérez
Track 3: Computer Vision Robotics and Intelligent Systems
Thu 14 Jan 2021 at 16:00 in session PS T3.9

Auto-TLDR; Adaptive Visual Semantic Specialized Network for Video Captioning

As an essential high-level task of video understanding topic, automatically describing a video with natural language has recently gained attention as a fundamental challenge in computer vision. Previous models for video captioning have several limitations, such as the existence of gaps in current semantic representations and the inexpressibility of the generated captions. To deal with these limitations, in this paper, we present a new architecture that we callAttentive Visual Semantic Specialized Network(AVSSN), which is an encoder-decoder model based on our Adaptive Attention Gate and Specialized LSTM layers. This architecture can selectively decide when to use visual or semantic information into the text generation process. The adaptive gate makes the decoder to automatically select the relevant information for providing a better temporal state representation than the existing decoders. Besides, the model is capable of learning to improve the expressiveness of generated captions attending to their length, using a sentence-length-related loss function. We evaluate the effectiveness of the proposed approach on the Microsoft Video Description(MSVD) and the Microsoft Research Video-to-Text (MSR-VTT) datasets, achieving state-of-the-art performance with several popular evaluation metrics: BLEU-4, METEOR, CIDEr, and ROUGE_L.

## Similar papers

#### Context Visual Information-Based Deliberation Network for Video Captioning

Min Lu, Xueyong Li, Caihua Liu
Track 3: Computer Vision Robotics and Intelligent Systems
Thu 14 Jan 2021 at 16:00 in session PS T3.9

Auto-TLDR; Context visual information-based deliberation network for video captioning

Video captioning is to automatically and accurately generate a textual description for a video. The typical methods following the encoder-decoder architecture directly utilized hidden states to predict words. Nevertheless, these methods did not amend the inaccurate hidden states before feeding those states into word prediction. This led to a cascade of errors on generating word by word. In this paper, the context visual information-based deliberation network is proposed, abbreviated as CVI-DelNet. Its key idea is to introduce the deliberator into the encoder-decoder framework. The encoder-decoder firstly generates a raw hidden state sequence. Unlike the existing methods, the raw hidden state is no more directly used for word prediction but is fed into the deliberator to generate the refined hidden state. The words are then predicted according to the refined hidden states and the contextual visual features. Results on two datasets shows that the proposed method significantly outperforms the baselines.

#### Visual Oriented Encoder: Integrating Multimodal and Multi-Scale Contexts for Video Captioning

Bang Yang, Yuexian Zou
Track 3: Computer Vision Robotics and Intelligent Systems
Thu 14 Jan 2021 at 14:00 in session PS T3.8

Auto-TLDR; Visual Oriented Encoder for Video Captioning

Video captioning is a challenging task which aims at automatically generating a natural language description of a given video. Recent researches have shown that exploiting the intrinsic multi-modalities of videos significantly promotes captioning performance. However, how to integrate multi-modalities to generate effective semantic representations for video captioning is still an open issue. Some researchers proposed to learn multimodal features in parallel during the encoding stage. The downside of these methods lies in the neglect of the interaction among multi-modalities and their rich contextual information. In this study, inspired by the fact that visual contents are generally more important for comprehending videos, we propose a novel Visual Oriented Encoder (VOE) to integrate multimodal features in an interactive manner. Specifically, VOE is designed as a hierarchical structure, where bottom layers are utilized to extract multi-scale contexts from auxiliary modalities while the top layer is exploited to generate joint representations by considering both visual and contextual information. Following the encoder-decoder framework, we systematically develop a VOE-LSTM model and evaluate it on two mainstream benchmarks: MSVD and MSR-VTT. Experimental results show that the proposed VOE surpasses conventional encoders and our VOE-LSTM model achieves competitive results compared with state-of-the-art approaches.

#### Enriching Video Captions with Contextual Text

Philipp Rimle, Pelin Dogan, Markus Gross
Track 3: Computer Vision Robotics and Intelligent Systems
Fri 15 Jan 2021 at 16:00 in session PS T3.11

Auto-TLDR; Contextualized Video Captioning Using Contextual Text

Understanding video content and generating caption with context is an important and challenging task. Unlike prior methods that typically attempt to generate generic video captions without context, our architecture contextualizes captioning by infusing extracted information from relevant text data. We propose an end-to-end sequence-to-sequence model which generates video captions based on visual input, and mines relevant knowledge such as names and locations from contextual text. In contrast to previous approaches, we do not preprocess the text further, and let the model directly learn to attend over it. Guided by the visual input, the model is able to copy words from the contextual text via a pointer-generator network, allowing to produce more specific video captions. We show competitive performance on the News Video Dataset and, through ablation studies, validate the efficacy of contextual video captioning as well as individual design choices in our model architecture.

#### Text Synopsis Generation for Egocentric Videos

Aidean Sharghi, Niels Lobo, Mubarak Shah
Track 3: Computer Vision Robotics and Intelligent Systems
Tue 12 Jan 2021 at 14:00 in session OS T3.1

Auto-TLDR; Egocentric Video Summarization Using Multi-task Learning for End-to-End Learning

Mass utilization of body-worn cameras has led to a huge corpus of available egocentric video. Existing video summarization algorithms can accelerate browsing such videos by selecting (visually) interesting shots from them. Nonetheless, since the system user still has to watch the summary videos, browsing large video databases remain a challenge. Hence, in this work, we propose to generate a textual synopsis, consisting of a few sentences describing the most important events in a long egocentric videos. Users can read the short text to gain insight about the video, and more importantly, efficiently search through the content of a large video database using text queries. Since egocentric videos are long and contain many activities and events, using video-to-text algorithms results in thousands of descriptions, many of which are incorrect. Therefore, we propose a multi-task learning scheme to simultaneously generate descriptions for video segments and summarize the resulting descriptions in an end-to-end fashion. We Input a set of video shots and the network generates a text description for each shot. Next, visual-language content matching unit that is trained with a weakly supervised objective, identifies the correct descriptions. Finally, the last component of our network, called purport network, evaluates the descriptions all together to select the ones containing crucial information. Out of thousands of descriptions generated for the video, a few informative sentences are returned to the user. We validate our framework on the challenging UT Egocentric video dataset, where each video is between 3 to 5 hours long, associated with over 3000 textual descriptions on average. The generated textual summaries, including only 5 percent (or less) of the generated descriptions, are compared to groundtruth summaries in text domain using well-established metrics in natural language processing.

#### PIN: A Novel Parallel Interactive Network for Spoken Language Understanding

Peilin Zhou, Zhiqi Huang, Fenglin Liu, Yuexian Zou
Track 2: Biometrics, Human Analysis and Behavior Understanding
Wed 13 Jan 2021 at 12:00 in session PS T2.2

Auto-TLDR; Parallel Interactive Network for Spoken Language Understanding

Spoken Language Understanding (SLU) is an essential part of the spoken dialogue system, which typically consists of intent detection (ID) and slot filling (SF) tasks. Recently, recurrent neural networks (RNNs) based methods achieved the state-of-the-art for SLU. It is noted that, in the existing RNN-based approaches, ID and SF tasks are often jointly modeled to utilize the correlation information between them. However, we noted that, so far, the efforts to obtain better performance by supporting bidirectional and explicit information exchange between ID and SF are not well studied. In addition, few studies attempt to capture the local context information to enhance the performance of SF. Motivated by these findings, in this paper, Parallel Interactive Network (PIN) is proposed to model the mutual guidance between ID and SF. Specifically, given an utterance, a Gaussian self-attentive encoder is introduced to generate the context-aware feature embedding of the utterance which is able to capture local context information. Taking the feature embedding of the utterance, Slot2Intent module and Intent2Slot module are developed to capture the bidirectional information flow for ID and SF tasks. Finally, a cooperation mechanism is constructed to fuse the information obtained from Slot2Intent and Intent2Slot modules to further reduce the prediction bias. The experiments on two benchmark datasets, i.e., SNIPS and ATIS, demonstrate the effectiveness of our approach, which achieves a competitive result with state-of-the-art models. More encouragingly, by using the feature embedding of the utterance generated by the pre-trained language model BERT, our method achieves the state-of-the-art among all comparison approaches.

#### A Novel Actor Dual-Critic Model for Remote Sensing Image Captioning

Ruchika Chavhan, Biplab Banerjee, Xiao Xiang Zhu, Subhasis Chaudhuri
Track 3: Computer Vision Robotics and Intelligent Systems
Wed 13 Jan 2021 at 16:30 in session PS T3.5

Auto-TLDR; Actor Dual-Critic Training for Remote Sensing Image Captioning Using Deep Reinforcement Learning

We deal with the problem of generating textual captions from optical remote sensing (RS) images using the notion of deep reinforcement learning. Due to the high inter-class similarity in reference sentences describing remote sensing data, jointly encoding the sentences and images encourages prediction of captions that are semantically more precise than the ground truth in many cases. To this end, we introduce an Actor Dual-Critic training strategy where a second critic model is deployed in the form of an encoder-decoder RNN to encode the latent information corresponding to the original and generated captions. While all actor-critic methods use an actor to predict sentences for an image and a critic to provide rewards, our proposed encoder-decoder RNN guarantees high-level comprehension of images by sentence-to-image translation. We observe that the proposed model generates sentences on the test data highly similar to the ground truth and is successful in generating even better captions in many critical cases. Extensive experiments on the benchmark Remote Sensing Image Captioning Dataset (RSICD) and the UCM-captions dataset confirm the superiority of the proposed approach in comparison to the previous state-of-the-art where we obtain a gain of sharp increments in both the ROUGE-L and CIDEr measures.

#### Multi-Scale 2D Representation Learning for Weakly-Supervised Moment Retrieval

Ding Li, Rui Wu, Zhizhong Zhang, Yongqiang Tang, Wensheng Zhang
Track 3: Computer Vision Robotics and Intelligent Systems
Thu 14 Jan 2021 at 16:00 in session PS T3.9

Auto-TLDR; Multi-scale 2D Representation Learning for Weakly Supervised Video Moment Retrieval

Video moment retrieval aims to search the moment most relevant to a given language query. However, most existing methods in this community often require temporal boundary annotations which are expensive and time-consuming to label. Hence weakly supervised methods have been put forward recently by only using coarse video-level label. Despite effectiveness, these methods usually process moment candidates independently, while ignoring a critical issue that the natural temporal dependencies between candidates in different temporal scales. To cope with this issue, we propose a Multi-scale 2D Representation Learning method for weakly supervised video moment retrieval. Specifically, we first construct a two-dimensional map for each temporal scale to capture the temporal dependencies between candidates. Two dimensions in this map indicate the start and end time points of these candidates. Then, we select top-K candidates from each scale-varied map with a learnable convolutional neural network. With a newly designed Moments Evaluation Module, we obtain the alignment scores of the selected candidates. At last, the similarity between captions and language query is served as supervision for further training the candidates' selector. Experiments on two benchmark datasets Charades-STA and ActivityNet Captions demonstrate that our approach achieves superior performance to state-of-the-art results.

#### MAGNet: Multi-Region Attention-Assisted Grounding of Natural Language Queries at Phrase Level

Amar Shrestha, Krittaphat Pugdeethosapol, Haowen Fang, Qinru Qiu
Track 3: Computer Vision Robotics and Intelligent Systems
Wed 13 Jan 2021 at 16:30 in session PS T3.5

Grounding free-form textual queries necessitates an understanding of these textual phrases and its relation to the visual cues to reliably reason about the described locations. Spatial attention networks are known to learn this relationship and focus its gaze on salient objects in the image. Thus, we propose to utilize spatial attention networks for image-level visual-textual fusion preserving local (word) and global (phrase) information to refine region proposals with an in-network Region Proposal Network (RPN) and detect single or multiple regions for a phrase query. We focus only on the phrase query - ground truth pair (referring expression) for a model independent of the constraints of the datasets i.e. additional attributes, context etc. For such referring expression dataset ReferIt game, our Multi- region Attention-assisted Grounding network (MAGNet) achieves over 12% improvement over the state-of-the-art. Without the con- text from image captions and attribute information in Flickr30k Entities, we still achieve competitive results compared to the state- of-the-art.

#### Dual Path Multi-Modal High-Order Features for Textual Content Based Visual Question Answering

Yanan Li, Yuetan Lin, Hongrui Zhao, Donghui Wang
Track 3: Computer Vision Robotics and Intelligent Systems
Fri 15 Jan 2021 at 13:00 in session OS T3.6

Auto-TLDR; TextVQA: An End-to-End Visual Question Answering Model for Text-Based VQA

As a typical cross-modal problem, visual question answering (VQA) has received increasing attention from the communities of computer vision and natural language processing. Reading and reasoning about texts and visual contents in the images is a burgeoning and important research topic in VQA, especially for the visually impaired assistance applications. Given an image, it aims to predict an answer to a provided natural language question closely related to its textual contents. In this paper, we propose a novel end-to-end textual content based VQA model, which grounds question answering both on the visual and textual information. After encoding the image, question and recognized text words, it uses multi-modal factorized high-order modules and the attention mechanism to fuse question-image and question-text features respectively. The complex correlations among different features can be captured efficiently. To ensure the model's extendibility, it embeds candidate answers and recognized texts in a semantic embedding space and adopts semantic embedding based classifier to perform answer prediction. Extensive experiments on the newly proposed benchmark TextVQA demonstrate that the proposed model can achieve promising results.

#### Tackling Contradiction Detection in German Using Machine Translation and End-To-End Recurrent Neural Networks

Maren Pielka, Rafet Sifa, Lars Patrick Hillebrand, David Biesner, Rajkumar Ramamurthy, Anna Ladi, Christian Bauckhage
Track 4: Document and Media Analysis
Tue 12 Jan 2021 at 17:00 in session PS T4.1

Auto-TLDR; Contradiction Detection in Natural Language Inference using Recurrent Neural Networks

Natural Language Inference, and specifically Contradiction Detection, is still an unexplored topic with respect to German text. In this paper, we apply Recurrent Neural Network (RNN) methods to learn contradiction-specific sentence embeddings. Our data set for evaluation is a machine-translated version of the Stanford Natural Language Inference (SNLI) corpus. The results are compared to a baseline using unsupervised vectorization techniques, namely tf-idf and Flair, as well as state-of-the art transformer-based (MBERT) methods. We find that the end-to-end models outperform the models trained on unsupervised embeddings, which makes them the better choice in an empirical use case. The RNN methods also perform superior to MBERT on the translated data set.

#### Transformer Reasoning Network for Image-Text Matching and Retrieval

Nicola Messina, Fabrizio Falchi, Andrea Esuli, Giuseppe Amato
Track 4: Document and Media Analysis
Tue 12 Jan 2021 at 17:00 in session PS T4.1

Auto-TLDR; A Transformer Encoder Reasoning Network for Image-Text Matching in Large-Scale Information Retrieval

Image-text matching is an interesting and fascinating task in modern AI research. Despite the evolution of deep-learning-based image and text processing systems, multi-modal matching remains a challenging problem. In this work, we consider the problem of accurate image-text matching for the task of multi-modal large-scale information retrieval. State-of-the-art results in image-text matching are achieved by inter-playing image and text features from the two different processing pipelines, usually using mutual attention mechanisms. However, this invalidates any chance to extract separate visual and textual features needed for later indexing steps in large-scale retrieval systems. In this regard, we introduce the Transformer Encoder Reasoning Network (TERN), an architecture built upon one of the modern relationship-aware self-attentive architectures, the Transformer Encoder (TE). This architecture is able to separately reason on the two different modalities and to enforce a final common abstract concept space by sharing the weights of the deeper transformer layers. Thanks to this design, the implemented network is able to produce compact and very rich visual and textual features available for the successive indexing step. Experiments are conducted on the MS-COCO dataset, and we evaluate the results using a discounted cumulative gain metric with relevance computed exploiting caption similarities, in order to assess possibly non-exact but relevant search results. We demonstrate that on this metric we are able to achieve state-of-the-art results in the image retrieval task. Our code is freely available at https://github.com/mesnico/TERN.

#### Context Matters: Self-Attention for Sign Language Recognition

Fares Ben Slimane, Mohamed Bouguessa
Track 1: Artificial Intelligence, Machine Learning for Pattern Analysis
Tue 12 Jan 2021 at 15:00 in session PS T1.2

Auto-TLDR; Attentional Network for Continuous Sign Language Recognition

This paper proposes an attentional network for the task of Continuous Sign Language Recognition. The proposed approach exploits co-independent streams of data to model the sign language modalities. These different channels of information can share a complex temporal structure between each other. For that reason, we apply attention to synchronize and help capture entangled dependencies between the different sign language components. Even though Sign Language is multi-channel, handshapes represent the central entities in sign interpretation. Seeing handshapes in their correct context defines the meaning of a sign. Taking that into account, we utilize the attention mechanism to efficiently aggregate the hand features with their appropriate Spatio-temporal context for better sign recognition. We found that by doing so the model is able to identify the essential Sign Language components that revolve around the dominant hand and the face areas. We test our model on the benchmark dataset RWTH-PHOENIX-Weather 2014, yielding competitive results.

#### Integrating Historical States and Co-Attention Mechanism for Visual Dialog

Tianling Jiang, Yi Ji, Chunping Liu
Track 3: Computer Vision Robotics and Intelligent Systems
Thu 14 Jan 2021 at 14:00 in session PS T3.8

Auto-TLDR; Integrating Historical States and Co-attention for Visual Dialog

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.

#### A Novel Attention-Based Aggregation Function to Combine Vision and Language

Matteo Stefanini, Marcella Cornia, Lorenzo Baraldi, Rita Cucchiara
Track 3: Computer Vision Robotics and Intelligent Systems
Wed 13 Jan 2021 at 16:30 in session PS T3.6

Auto-TLDR; Fully-Attentive Reduction for Vision and Language

The joint understanding of vision and language has been recently gaining a lot of attention in both the Computer Vision and Natural Language Processing communities, with the emergence of tasks such as image captioning, image-text matching, and visual question answering. As both images and text can be encoded as sets or sequences of elements - like regions and words - proper reduction functions are needed to transform a set of encoded elements into a single response, like a classification or similarity score. In this paper, we propose a novel fully-attentive reduction method for vision and language. Specifically, our approach computes a set of scores for each element of each modality employing a novel variant of cross-attention, and performs a learnable and cross-modal reduction, which can be used for both classification and ranking. We test our approach on image-text matching and visual question answering, building fair comparisons with other reduction choices, on both COCO and VQA 2.0 datasets. Experimentally, we demonstrate that our approach leads to a performance increase on both tasks. Further, we conduct ablation studies to validate the role of each component of the approach.

#### Cross-Lingual Text Image Recognition Via Multi-Task Sequence to Sequence Learning

Zhuo Chen, Fei Yin, Xu-Yao Zhang, Qing Yang, Cheng-Lin Liu
Track 4: Document and Media Analysis
Wed 13 Jan 2021 at 12:00 in session PS T4.2

Auto-TLDR; Cross-Lingual Text Image Recognition with Multi-task Learning

This paper considers recognizing texts shown in a source language and translating into a target language, without generating the intermediate source language text image recognition results. We call this problem Cross-Lingual Text Image Recognition (CLTIR). To solve this problem, we propose a multi-task system containing a main task of CLTIR and an auxiliary task of Mono-Lingual Text Image Recognition (MLTIR) simultaneously. Two different sequence to sequence learning methods, a convolution based attention model and a BLSTM model with CTC, are adopted for these tasks respectively. We evaluate the system on a newly collected Chinese-English bilingual movie subtitle image dataset. Experimental results demonstrate the multi-task learning framework performs superiorly in both languages.

#### GCNs-Based Context-Aware Short Text Similarity Model

Xiaoqi Sun
Track 1: Artificial Intelligence, Machine Learning for Pattern Analysis
Tue 12 Jan 2021 at 15:00 in session PS T1.1

Auto-TLDR; Context-Aware Graph Convolutional Network for Text Similarity

Semantic textual similarity is a fundamental task in text mining and natural language processing (NLP), which has profound research value. The essential step for text similarity is text representation learning. Recently, researches have explored the graph convolutional network (GCN) techniques on text representation, since GCN does well in handling complex structures and preserving syntactic information. However, current GCN models are usually limited to very shallow layers due to the vanishing gradient problem, which cannot capture non-local dependency information of sentences. In this paper, we propose a GCNs-based context-aware (GCSTS) model that applies iterated GCN blocks to train deeper GCNs. Recurrently employing the same GCN block prevents over-fitting and provides broad effective input width. Combined with dense connections, GCSTS can be trained more deeply. Besides, we use dynamic graph structures in the block, which further extend the receptive field of each vertex in graphs, learning better sentence representations. Experiments show that our model outperforms existing models on several text similarity datasets, while also verify that GCNs-based text representation models can be trained in a deeper manner, rather than being trained in two or three layers.

#### MA-LSTM: A Multi-Attention Based LSTM for Complex Pattern Extraction

Jingjie Guo, Kelang Tian, Kejiang Ye, Cheng-Zhong Xu
Track 1: Artificial Intelligence, Machine Learning for Pattern Analysis
Wed 13 Jan 2021 at 12:00 in session PS T1.3

Auto-TLDR; MA-LSTM: Multiple Attention based recurrent neural network for forget gate

With the improvement of data, computing powerand algorithms, deep learning has achieved rapid developmentand showing excellent performance. Recently, many deep learn-ing models are proposed to solve the problems in different areas.A recurrent neural network (RNN) is a class of artificial neuralnetworks where connections between nodes form a directedgraph along a temporal sequence. This allows it to exhibittemporal dynamic behavior, which makes it applicable to taskssuch as handwriting recognition or speech recognition. How-ever, the RNN relies heavily on the automatic learning abilityto update parameters which concentrate on the data flow butseldom considers the feature extraction capability of the gatemechanism. In this paper, we propose a novel architecture tobuild the forget gate which is generated by multiple bases.Instead of using the traditional single-layer fully-connectednetwork, we use a Multiple Attention (MA) based network togenerate the forget gate which refines the optimization spaceof gate function and improve the granularity of the recurrentneural network to approximate the map in the ground truth.Credit to the MA structure on the gate mechanism. Our modelhas a better feature extraction capability than other knownmodels. MA-LSTM is an alternative module which can directly replace the recurrent neural network and has achieved good performance in many areas that people are concerned about.

#### Explore and Explain: Self-Supervised Navigation and Recounting

Roberto Bigazzi, Federico Landi, Marcella Cornia, Silvia Cascianelli, Lorenzo Baraldi, Rita Cucchiara
Track 3: Computer Vision Robotics and Intelligent Systems
Wed 13 Jan 2021 at 14:00 in session OS T3.2

Auto-TLDR; Exploring a Photorealistic Environment for Explanation and Navigation

Embodied AI has been recently gaining attention as it aims to foster the development of autonomous and intelligent agents. In this paper, we devise a novel embodied setting in which an agent needs to explore a previously unknown environment while recounting what it sees during the path. In this context, the agent needs to navigate the environment driven by an exploration goal, select proper moments for description, and output natural language descriptions of relevant objects and scenes. Our model integrates a novel self-supervised exploration module with penalty, and a fully-attentive captioning model for explanation. Also, we investigate different policies for selecting proper moments for explanation, driven by information coming from both the environment and the navigation. Experiments are conducted on photorealistic environments from the Matterport3D dataset and investigate the navigation and explanation capabilities of the agent as well as the role of their interactions.

#### Multi-Modal Contextual Graph Neural Network for Text Visual Question Answering

Yaoyuan Liang, Xin Wang, Xuguang Duan, Wenwu Zhu
Track 4: Document and Media Analysis
Thu 14 Jan 2021 at 12:00 in session PS T4.3

Auto-TLDR; Multi-modal Contextual Graph Neural Network for Text Visual Question Answering

Text visual question answering (TextVQA) targets at answering the question related to texts appearing in the given images, posing more challenges than VQA by requiring a deeper recognition and understanding of various shapes of human-readable scene texts as well as their meanings in different contexts. Existing works on TextVQA suffer from two weaknesses: i) scene texts and non-textual objects are processed separately and independently without considering their mutual interactions during the question understanding and answering process, ii) scene texts are encoded only through word embeddings without taking the corresponding visual appearance features as well as their potential relationships with other non-textual objects in the images into account. To overcome the weakness of exiting works, we propose a novel multi-modal contextual graph neural network (MCG) model for TextVQA. The proposed MCG model can capture the relationships between visual features of scene texts and non-textual objects in the given images as well as utilize richer sources of multi-modal features to improve the model performance. In particular, we encode the scene texts into richer features containing textual, visual and positional features, then model the visual relations between scene texts and non-textual objects through a contextual graph neural network. Our extensive experiments on real-world dataset demonstrate the advantages of the proposed MCG model over baseline approaches.

#### Extracting Action Hierarchies from Action Labels and their Use in Deep Action Recognition

Track 3: Computer Vision Robotics and Intelligent Systems
Tue 12 Jan 2021 at 15:00 in session PS T3.1

Auto-TLDR; Exploiting the Information Content of Language Label Associations for Human Action Recognition

Human activity recognition is a fundamental and challenging task in computer vision. Its solution can support multiple and diverse applications in areas including but not limited to smart homes, surveillance, daily living assistance, Human-Robot Collaboration (HRC), etc. In realistic conditions, the complexity of human activities ranges from simple coarse actions, such as siting or standing up, to more complex activities that consist of multiple actions with subtle variations in appearance and motion patterns. A large variety of existing datasets target specific action classes, with some of them being coarse and others being fine-grained. In all of them, a description of the action and its complexity is manifested in the action label sentence. As the action/activity complexity increases, so is the label sentence size and the amount of action-related semantic information contained in this description. In this paper, we propose an approach to exploit the information content of these action labels to formulate a coarse-to-fine action hierarchy based on linguistic label associations, and investigate the potential benefits and drawbacks. Moreover, in a series of quantitative and qualitative experiments, we show that the exploitation of this hierarchical organization of action classes in different levels of granularity improves the learning speed and overall performance of a range of baseline and mid-range deep architectures for human action recognition (HAR).

#### Cross-Supervised Joint-Event-Extraction with Heterogeneous Information Networks

Yue Wang, Zhuo Xu, Yao Wan, Lu Bai, Lixin Cui, Qian Zhao, Edwin Hancock, Philip Yu
Track 4: Document and Media Analysis
Wed 13 Jan 2021 at 12:00 in session PS T4.2

Auto-TLDR; Joint-Event-extraction from Unstructured corpora using Structural Information Network

Joint-event-extraction, which extracts structural information (i.e., entities or triggers of events) from unstructured real-world corpora, has attracted more and more research attention in natural language processing. \revised{Most existing works do not fully address the sparse co-occurred relationships between entities and triggers. This exacerbates the error-propagation problem} which may degrade the extraction performance. To mitigate this issue, we first define the joint-event-extraction as a sequence-to-sequence labeling task with a tag set which is composed of tags of triggers and entities. Then, to incorporate the missing information in the aforementioned co-occurred relationships, we propose a \underline{C}ross-\underline{S}upervised \underline{M}echanism (CSM) to alternately supervise the extraction of either triggers or entities based on the type distribution of each other. Moreover, since the connected entities and triggers naturally form a heterogeneous information network (HIN), we leverage the latent pattern along meta-paths for a given corpus to further improve the performance of our proposed method. To verify the effectiveness of our proposed method, we conduct extensive experiments on real-world datasets as well as compare our method with state-of-the-art methods. Empirical results and analysis show that our approach outperforms the state-of-the-art methods in both entity and trigger extraction.

#### Trajectory-User Link with Attention Recurrent Networks

Tao Sun, Yongjun Xu, Fei Wang, Lin Wu, 塘文 钱, Zezhi Shao
Track 1: Artificial Intelligence, Machine Learning for Pattern Analysis
Wed 13 Jan 2021 at 14:00 in session PS T1.6

Auto-TLDR; TULAR: Trajectory-User Link with Attention Recurrent Neural Networks

The prevalent adoptions of GPS-enabled devices have witnessed an explosion of various location-based services which produces a huge amount of trajectories monitoring the individuals' movements. In this paper, we tackle Trajectory-User Link (TUL) problem, which identifies humans' movement patterns and links trajectories to the users who generated them. Existing solutions on TUL problem employ recurrent neural networks and variational autoencoder methods, which face the bottlenecks in the case of excessively long trajectories and fragmentary users' movements. However, these are common characteristics of trajectory data in reality, leading to performance degradation of the existing models. In this paper, we propose an end-to-end attention recurrent neural learning framework, called TULAR (Trajectory-User Link with Attention Recurrent Networks), which focus on selected parts of the source trajectories when linking. TULAR introduce the Trajectory Semantic Vector (TSV) via unsupervised location representation learning and recurrent neural networks, by which to reckon the weight of parts of source trajectory. Further, we employ three attention scores for the weight measurements. Experiments are conducted on two real world datasets and compared with several existing methods, and the results show that TULAR yields a new state-of-the-art performance. Source code is public available at GitHub: https://github.com/taos123/TULAR.

#### Flow-Guided Spatial Attention Tracking for Egocentric Activity Recognition

Tianshan Liu, Kin-Man Lam
Track 2: Biometrics, Human Analysis and Behavior Understanding
Thu 14 Jan 2021 at 12:00 in session PS T2.4

Auto-TLDR; flow-guided spatial attention tracking for egocentric activity recognition

The popularity of wearable cameras has opened up a new dimension for egocentric activity recognition. While some methods introduce attention mechanisms into deep learning networks to capture fine-grained hand-object interactions, they often neglect exploring the spatio-temporal relationships. Generating spatial attention, without adequately exploiting temporal consistency, will result in potentially sub-optimal performance in the video-based task. In this paper, we propose a flow-guided spatial attention tracking (F-SAT) module, which is based on enhancing motion patterns and inter-frame information, to highlight the discriminative features from regions of interest across a video sequence. A new form of input, namely the optical-flow volume, is presented to provide informative cues from moving parts for spatial attention tracking. The proposed F-SAT module is deployed to a two-branch-based deep architecture, which fuses complementary information for egocentric activity recognition. Experimental results on three egocentric activity benchmarks show that the proposed method achieves state-of-the-art performance.

#### Efficient Sentence Embedding Via Semantic Subspace Analysis

Bin Wang, Fenxiao Chen, Yun Cheng Wang, C.-C. Jay Kuo
Track 4: Document and Media Analysis
Tue 12 Jan 2021 at 17:00 in session PS T4.1

Auto-TLDR; S3E: Semantic Subspace Sentence Embedding

A novel sentence embedding method built upon semantic subspace analysis, called semantic subspace sentence embedding (S3E), is proposed in this work. Given the fact that word embeddings can capture semantic relationship while semantically similar words tend to form semantic groups in a high-dimensional embedding space, we develop a sentence representation scheme by analyzing semantic subspaces of its constituent words. Specifically, we construct a sentence model from two aspects. First, we represent words that lie in the same semantic group using the intra-group descriptor. Second, we characterize the interaction between multiple semantic groups with the inter-group descriptor. The proposed S3E method is evaluated on both textual similarity tasks and supervised tasks. Experimental results show that it offers comparable or better performance than the state-of-the-art. The complexity of our S3E method is also much lower than other parameterized models.

#### PICK: Processing Key Information Extraction from Documents Using Improved Graph Learning-Convolutional Networks

Wenwen Yu, Ning Lu, Xianbiao Qi, Ping Gong, Rong Xiao
Track 4: Document and Media Analysis
Wed 13 Jan 2021 at 12:00 in session PS T4.2

Auto-TLDR; PICK: A Graph Learning Framework for Key Information Extraction from Documents

Computer vision with state-of-the-art deep learning models have achieved huge success in the field of Optical Character Recognition (OCR) including text detection and recognition tasks recently. However, Key Information Extraction (KIE) from documents as the downstream task of OCR, having a large number of use scenarios in real-world, remains a challenge because documents not only have textual features extracting from OCR systems but also have semantic visual features that are not fully exploited and play a critical role in KIE. Too little work has been devoted to efficiently make full use of both textual and visual features of the documents. In this paper, we introduce PICK, a framework that is effective and robust in handling complex documents layout for KIE by combining graph learning with graph convolution operation, yielding a richer semantic representation containing the textual and visual features and global layout without ambiguity. Extensive experiments on real-world datasets have been conducted to show that our method outperforms baselines methods by significant margins.

#### Continuous Sign Language Recognition with Iterative Spatiotemporal Fine-Tuning

Kenessary Koishybay, Medet Mukushev, Anara Sandygulova
Track 1: Artificial Intelligence, Machine Learning for Pattern Analysis
Fri 15 Jan 2021 at 16:00 in session PS T1.16

Auto-TLDR; A Deep Neural Network for Continuous Sign Language Recognition with Iterative Gloss Recognition

This paper aims to develop a deep neural network for Continuous Sign Language Recognition (CSLR) with iterative Gloss Recognition (GR) fine-tuning. CSLR has been a popular research field in the last years and iterative optimization methods are well established. This paper introduces our proposed architecture involving Spatiotemporal feature-extraction model to segment useful gloss-unit" features and BiLSTM with CTC as a sequence model. Spatiotemporal Feature Extractor is used for both image features extraction and sequence length reduction. To this end, we compare different architectures for feature extraction and sequence model. In addition, we iteratively fine-tune feature extractor on gloss-unit video segments with alignments from the end2end model. During the iterative training, we use novel alignment correction technique, which is based on minimum transformations of Levenshtein distance. All the experiments were conducted on the RWTH-PHOENIX-Weather-2014 dataset.

#### Zero-Shot Text Classification with Semantically Extended Graph Convolutional Network

Tengfei Liu, Yongli Hu, Junbin Gao, Yanfeng Sun, Baocai Yin
Track 4: Document and Media Analysis
Wed 13 Jan 2021 at 12:00 in session PS T4.2

Auto-TLDR; Semantically Extended Graph Convolutional Network for Zero-shot Text Classification

As a challenging task of Natural Language Processing(NLP), zero-shot text classification has attracted more and more attention recently. It aims to detect classes that the model has never seen in the training set. For this purpose, a feasible way is to construct connection between the seen and unseen classes by semantic extension and classify the unseen classes by information propagation over the connection. Although many related zero-shot text classification methods have been exploited, how to realize semantic extension properly and propagate information effectively is far from solved. In this paper, we propose a novel zero-shot text classification method called Semantically Extended Graph Convolutional Network (SEGCN). In the proposed method, the semantic category knowledge from ConceptNet is utilized to semantic extension for linking seen classes to unseen classes and constructing a graph of all classes. Then, we build upon Graph Convolutional Network (GCN) for predicting the textual classifier for each category, which transfers the category knowledge by the convolution operators on the constructed graph and is trained in a semi-supervised manner using the samples of the seen classes. The experimental results on Dbpedia and 20newsgroup datasets show that our method outperforms the state of the art zero-shot text classification methods.

#### Global Feature Aggregation for Accident Anticipation

Mishal Fatima, Umar Karim Khan, Chong Min Kyung
Track 3: Computer Vision Robotics and Intelligent Systems
Tue 12 Jan 2021 at 14:00 in session OS T3.1

Auto-TLDR; Feature Aggregation for Predicting Accidents in Video Sequences

Anticipation of accidents ahead of time in autonomous and non-autonomous vehicles aids in accident avoidance. In order to recognize abnormal events such as traffic accidents in a video sequence, it is important that the network takes into account interactions of objects in a given frame. We propose a novel Feature Aggregation (FA) block that refines each object's features by computing a weighted sum of the features of all objects in a frame. We use FA block along with Long Short Term Memory (LSTM) network to anticipate accidents in the video sequences. We report mean Average Precision (mAP) and Average Time-to-Accident (ATTA) on Street Accident (SA) dataset. Our proposed method achieves the highest score for risk anticipation by predicting accidents 0.32 sec and 0.75 sec earlier compared to the best results with Adaptive Loss and dynamic parameter prediction based methods respectively.

#### SAT-Net: Self-Attention and Temporal Fusion for Facial Action Unit Detection

Zhihua Li, Zheng Zhang, Lijun Yin
Track 2: Biometrics, Human Analysis and Behavior Understanding
Tue 12 Jan 2021 at 17:00 in session PS T2.1

Auto-TLDR; Temporal Fusion and Self-Attention Network for Facial Action Unit Detection

Research on facial action unit detection has shown remarkable performances by using deep spatial learning models in recent years, however, it is far from reaching its full capacity in learning due to the lack of use of temporal information of AUs across time. Since the AU occurrence in one frame is highly likely related to previous frames in a temporal sequence, exploring temporal correlation of AUs across frames becomes a key motivation of this work. In this paper, we propose a novel temporal fusion and AU-supervised self-attention network (a so-called SAT-Net) to address the AU detection problem. First of all, we input the deep features of a sequence into a convolutional LSTM network and fuse the previous temporal information into the feature map of the last frame, and continue to learn the AU occurrence. Second, considering the AU detection problem is a multi-label classification problem that individual label depends only on certain facial areas, we propose a new self-learned attention mask by focusing the detection of each AU on parts of facial areas through the learning of individual attention mask for each AU, thus increasing the AU independence without the loss of any spatial relations. Our extensive experiments show that the proposed framework achieves better results of AU detection over the state-of-the-arts on two benchmark databases (BP4D and DISFA).

#### A CNN-RNN Framework for Image Annotation from Visual Cues and Social Network Metadata

Tobia Tesan, Pasquale Coscia, Lamberto Ballan
Track 3: Computer Vision Robotics and Intelligent Systems
Wed 13 Jan 2021 at 16:30 in session PS T3.6

Auto-TLDR; Context-Based Image Annotation with Multiple Semantic Embeddings and Recurrent Neural Networks

Images represent a commonly used form of visual communication among people. Nevertheless, image classification may be a challenging task when dealing with unclear or non-common images needing more context to be correctly annotated. Metadata accompanying images on social-media represent an ideal source of additional information for retrieving proper neighborhoods easing image annotation task. To this end, we blend visual features extracted from neighbors and their metadata to jointly leverage context and visual cues. Our models use multiple semantic embeddings to achieve the dual objective of being robust to vocabulary changes between train and test sets and decoupling the architecture from the low-level metadata representation. Convolutional and recurrent neural networks (CNNs-RNNs) are jointly adopted to infer similarity among neighbors and query images. We perform comprehensive experiments on the NUS-WIDE dataset showing that our models outperform state-of-the-art architectures based on images and metadata, and decrease both sensory and semantic gaps to better annotate images.

#### Reinforcement Learning with Dual Attention Guided Graph Convolution for Relation Extraction

Zhixin Li, Yaru Sun, Suqin Tang, Canlong Zhang, Huifang Ma
Track 4: Document and Media Analysis
Fri 15 Jan 2021 at 15:00 in session PS T1.13

Auto-TLDR; Dual Attention Graph Convolutional Network for Relation Extraction

To better learn the dependency relationship between nodes, we address the relationship extraction task by capturing rich contextual dependencies based on the attention mechanism, and using distributional reinforcement learning to generate optimal relation information representation. This method is called Dual Attention Graph Convolutional Network (DAGCN), to adaptively integrate local features with their global dependencies. Specifically, we append two types of attention modules on top of GCN, which model the semantic interdependencies in spatial and relational dimensions respectively. The position attention module selectively aggregates the feature at each position by a weighted sum of the features at all positions of nodes internal features. Meanwhile, the relation attention module selectively emphasizes interdependent node relations by integrating associated features among all nodes. We sum the outputs of the two attention modules and use reinforcement learning to predict the classification of nodes relationship to further improve feature representation which contributes to more precise extraction results. The results on the TACRED and SemEval datasets show that the model can obtain more useful information for relational extraction tasks, and achieve better performances on various evaluation indexes.

#### ConvMath : A Convolutional Sequence Network for Mathematical Expression Recognition

Zuoyu Yan, Xiaode Zhang, Liangcai Gao, Ke Yuan, Zhi Tang
Track 4: Document and Media Analysis
Wed 13 Jan 2021 at 12:00 in session PS T4.2

Auto-TLDR; Convolutional Sequence Modeling for Mathematical Expressions Recognition

Despite the recent advances in optical character recognition (OCR), mathematical expressions still face a great challenge to recognize due to their two-dimensional graphical layout. In this paper, we propose a convolutional sequence modeling network, ConvMath, which converts the mathematical expression description in an image into a LaTeX sequence in an end-to-end way. The network combines an image encoder for feature extraction and a convolutional decoder for sequence generation. Compared with other Long Short Term Memory(LSTM) based encoder-decoder models, ConvMath is entirely based on convolution, thus it is easy to perform parallel computation. Besides, the network adopts multi-layer attention mechanism in the decoder, which allows the model to align output symbols with source feature vectors automatically, and alleviates the problem of lacking coverage while training the model. The performance of ConvMath is evaluated on an open dataset named IM2LATEX-100K, including 103556 samples. The experimental results demonstrate that the proposed network achieves state-of-the-art accuracy and much better efficiency than previous methods.

#### Global Context-Based Network with Transformer for Image2latex

Nuo Pang, Chun Yang, Xiaobin Zhu, Jixuan Li, Xu-Cheng Yin
Track 4: Document and Media Analysis
Wed 13 Jan 2021 at 12:00 in session PS T4.2

Auto-TLDR; Image2latex with Global Context block and Transformer

Image2latex usually means converts mathematical formulas in images into latex markup. It is a very challenging job due to the complex two-dimensional structure, variant scales of input, and very long representation sequence. Many researchers use encoder-decoder based model to solve this task and achieved good results. However, these methods don't make full use of the structure and position information of the formula. %In this paper, we improve the encoder by employing Global Context block and Transformer. To solve this problem, we propose a global context-based network with transformer that can (1) learn a more powerful and robust intermediate representation via aggregating global features and (2) encode position information explicitly and (3) learn latent dependencies between symbols by using self-attention mechanism. The experimental results on the dataset IM2LATEX-100K demonstrate the effectiveness of our method.

#### 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.

#### Developing Motion Code Embedding for Action Recognition in Videos

Maxat Alibayev, David Andrea Paulius, Yu Sun
Track 3: Computer Vision Robotics and Intelligent Systems
Tue 12 Jan 2021 at 17:00 in session PS T3.3

Auto-TLDR; Motion Embedding via Motion Codes for Action Recognition

We propose a motion embedding strategy via the motion codes that is a vectorized representation of motions based on their salient mechanical attributes. We show that our motion codes can provide robust motion representation. We train a deep neural network model that learns to embed demonstration videos into motion codes. We integrate the extracted features from the motion embedding model into the current state-of-the-art action recognition model. The obtained model achieved higher accuracy than the baseline on a verb classification task from egocentric videos in EPIC-KITCHENS dataset.

#### Revisiting Sequence-To-Sequence Video Object Segmentation with Multi-Task Loss and Skip-Memory

Fatemeh Azimi, Benjamin Bischke, Sebastian Palacio, Federico Raue, Jörn Hees, Andreas Dengel
Track 3: Computer Vision Robotics and Intelligent Systems
Fri 15 Jan 2021 at 16:00 in session PS T3.11

Auto-TLDR; Sequence-to-Sequence Learning for Video Object Segmentation

Video Object Segmentation (VOS) is an active research area of the visual domain. One of its fundamental sub-tasks is semi-supervised / one-shot learning: given only the segmentation mask for the first frame, the task is to provide pixel-accurate masks for the object over the rest of the sequence. Despite much progress in the last years, we noticed that many of the existing approaches lose objects in longer sequences, especially when the object is small or briefly occluded. In this work, we build upon a sequence-to-sequence approach that employs an encoder-decoder architecture together with a memory module for exploiting the sequential data. We further improve this approach by proposing a model that manipulates multi-scale spatio-temporal information using memory-equipped skip connections. Furthermore, we incorporate an auxiliary task based on distance classification which greatly enhances the quality of edges in segmentation masks. We compare our approach to the state of the art and show considerable improvement in the contour accuracy metric and the overall segmentation accuracy.

#### KoreALBERT: Pretraining a Lite BERT Model for Korean Language Understanding

Hyunjae Lee, Jaewoong Yun, Bongkyu Hwang, Seongho Joe, Seungjai Min, Youngjune Gwon
Track 4: Document and Media Analysis
Wed 13 Jan 2021 at 12:00 in session PS T4.2

Auto-TLDR; KoreALBERT: A monolingual ALBERT model for Korean language understanding

Abstract—A Lite BERT (ALBERT) has been introduced to scale-up deep bidirectional representation learning for natural languages. Due to the lack of pretrained ALBERT models for Korean language, the best available practice is the multilingual model or resorting back to the any other BERT-based model. In this paper, we develop and pretrain KoreALBERT, a monolingual ALBERT model specifically for Korean language understanding. We introduce a new training objective, namely Word Order Prediction (WOP), and use alongside the existing MLM and SOP criteria to the same architecture and model parameters. Despite having significantly fewer model parameters (thus, quicker to train), our pretrained KoreALBERT outperforms its BERT counterpart on KorQuAD 1.0 benchmark for machine reading comprehension. Consistent with the empirical results in English by Lan et al., KoreALBERT seems to improve downstream task performance involving multi-sentence encoding for Korean language. The pretrained KoreALBERT is publicly available to encourage research and application development for Korean NLP.

#### Evaluation of BERT and ALBERT Sentence Embedding Performance on Downstream NLP Tasks

Hyunjin Choi, Judong Kim, Seongho Joe, Youngjune Gwon
Track 4: Document and Media Analysis
Wed 13 Jan 2021 at 12:00 in session PS T4.2

Auto-TLDR; Sentence Embedding Models for BERT and ALBERT: A Comparison and Evaluation

Contextualized representations from a pre-trained language model are central to achieve a high performance on downstream NLP task. The pre-trained BERT and A Lite BERT (ALBERT) models can be fine-tuned to give state-of-the-art results in sentence-pair regressions such as semantic textual similarity (STS) and natural language inference (NLI). Although BERT-based models yield the [CLS] token vector as a reasonable sentence embedding, the search for an optimal sentence embedding scheme remains an active research area in computational linguistics. This paper explores on sentence embedding models for BERT and ALBERT. In particular, we take a modified BERT network with siamese and triplet network structures called Sentence-BERT (SBERT) and replace BERT with ALBERT to create Sentence-ALBERT (SALBERT). We also experiment with an outer CNN sentence-embedding network for SBERT and SALBERT. We evaluate performances of all sentence-embedding models considered using the STS and NLI datasets. The empirical results indicate that our CNN architecture improves ALBERT models substantially more than BERT models for STS benchmark. Despite significantly fewer model parameters, ALBERT sentence embedding is highly competitive to BERT in downstream NLP evaluations.

#### PrivAttNet: Predicting Privacy Risks in Images Using Visual Attention

Chen Zhang, Thivya Kandappu, Vigneshwaran Subbaraju
Track 3: Computer Vision Robotics and Intelligent Systems
Thu 14 Jan 2021 at 16:00 in session PS T3.9

Auto-TLDR; PrivAttNet: A Visual Attention Based Approach for Privacy Sensitivity in Images

Visual privacy concerns associated with image sharing is a critical issue that need to be addressed to enable safe and lawful use of online social platforms. Users of social media platforms often suffer from no guidance in sharing sensitive images in public, and often face with social and legal consequences. Given the recent success of visual attention based deep learning methods in measuring abstract phenomena like image memorability, we are motivated to investigate whether visual attention based methods could be useful in measuring psycho-physical phenomena like "privacy sensitivity". In this paper we propose PrivAttNet -- a visual attention based approach, that can be trained end-to-end to estimate the privacy sensitivity of images without explicitly detecting objects and attributes present in the image. We show that our PrivAttNet model outperforms various SOTA and baseline strategies -- a 1.6 fold reduction in $L1-error$ over SOTA and 7%--10% improvement in Spearman-rank correlation between the predicted and ground truth sensitivity scores. Additionally, the attention maps from PrivAttNet are found to be useful in directing the users to the regions that are responsible for generating the privacy risk score.

#### VSR++: Improving Visual Semantic Reasoning for Fine-Grained Image-Text Matching

Hui Yuan, Yan Huang, Dongbo Zhang, Zerui Chen, Wenlong Cheng, Liang Wang
Track 3: Computer Vision Robotics and Intelligent Systems
Thu 14 Jan 2021 at 12:00 in session PS T3.7

Auto-TLDR; Improving Visual Semantic Reasoning for Fine-Grained Image-Text Matching

Image-text matching has made great progresses recently, but there still remains challenges in fine-grained matching. To deal with this problem, we propose an Improved Visual Semantic Reasoning model (VSR++), which jointly models 1) global alignment between images and texts and 2) local correspondence between regions and words in a unified framework. To exploit their complementary advantages, we also develop a suitable learning strategy to balance their relative importance. As a result, our model can distinguish image regions and text words in a fine-grained level, and thus achieves the current stateof-the-art performance on two benchmark datasets.

#### Multi-Stage Attention Based Visual Question Answering

Aakansha Mishra, Ashish Anand, Prithwijit Guha
Track 3: Computer Vision Robotics and Intelligent Systems
Wed 13 Jan 2021 at 16:30 in session PS T3.6

Auto-TLDR; Alternative Bi-directional Attention for Visual Question Answering

Recent developments in the field of Visual Question Answering (VQA) have witnessed promising improvements in performance through contributions in attention based networks. Most such approaches have focused on unidirectional attention that leverage over attention from textual domain (question) on visual space. These approaches mostly focused on learning high-quality attention in the visual space. In contrast, this work proposes an alternating bi-directional attention framework. First, a question to image attention helps to learn the robust visual space embedding, and second, an image to question attention helps to improve the question embedding. This attention mechanism is realized in an alternating fashion i.e. question-to-image followed by image-to-question and is repeated for maximizing performance. We believe that this process of alternating attention generation helps both the modalities and leads to better representations for the VQA task. This proposal is benchmark on TDIUC dataset and against state-of-art approaches. Our ablation analysis shows that alternate attention is the key to achieve high performance in VQA.

#### AG-GAN: An Attentive Group-Aware GAN for Pedestrian Trajectory Prediction

Yue Song, Niccolò Bisagno, Syed Zohaib Hassan, Nicola Conci
Track 2: Biometrics, Human Analysis and Behavior Understanding
Wed 13 Jan 2021 at 14:00 in session PS T2.3

Auto-TLDR; An attentive group-aware GAN for motion prediction in crowded scenarios

Understanding human behaviors in crowded scenarios requires analyzing not only the position of the subjects in space, but also the scene context. Existing approaches mostly rely on the motion history of each pedestrian and model the interactions among people by considering the entire surrounding neighborhood. In our approach, we address the problem of motion prediction by applying coherent group clustering and a global attention mechanism on the LSTM-based Generative Adversarial Networks (GANs). The proposed model consists of an attentive group-aware GAN that observes the agents' past motion and predicts future paths, using (i) a group pooling module to model neighborhood interaction, and (ii) an attention module to specifically focus on hidden states. The experimental results demonstrate that our proposal outperforms state-of-the-art models on common benchmark datasets, and is able to generate socially-acceptable trajectories.

#### A Grid-Based Representation for Human Action Recognition

Soufiane Lamghari, Guillaume-Alexandre Bilodeau, Nicolas Saunier
Track 2: Biometrics, Human Analysis and Behavior Understanding
Tue 12 Jan 2021 at 17:00 in session PS T2.1

Auto-TLDR; GRAR: Grid-based Representation for Action Recognition in Videos

Human action recognition (HAR) in videos is a fundamental research topic in computer vision. It consists mainly in understanding actions performed by humans based on a sequence of visual observations. In recent years, HAR have witnessed significant progress, especially with the emergence of deep learning models. However, most of existing approaches for action recognition rely on information that is not always relevant for the task, and are limited in the way they fuse temporal information. In this paper, we propose a novel method for human action recognition that encodes efficiently the most discriminative appearance information of an action with explicit attention on representative pose features, into a new compact grid representation. Our GRAR (Grid-based Representation for Action Recognition) method is tested on several benchmark datasets that demonstrate that our model can accurately recognize human actions, despite intra-class appearance variations and occlusion challenges.

#### ActionSpotter: Deep Reinforcement Learning Framework for Temporal Action Spotting in Videos

Guillaume Vaudaux-Ruth, Adrien Chan-Hon-Tong, Catherine Achard
Track 3: Computer Vision Robotics and Intelligent Systems
Tue 12 Jan 2021 at 17:00 in session PS T3.3

Auto-TLDR; ActionSpotter: A Reinforcement Learning Algorithm for Action Spotting in Video

Action spotting has recently been proposed as an alternative to action detection and key frame extraction. However, the current state-of-the-art method of action spotting requires an expensive ground truth composed of the search sequences employed by human annotators spotting actions - a critical limitation. In this article, we propose to use a reinforcement learning algorithm to perform efficient action spotting using only the temporal segments from the action detection annotations, thus opening an interesting solution for video understanding. Experiments performed on THUMOS14 and ActivityNet datasets show that the proposed method, named ActionSpotter, leads to good results and outperforms state-of-the-art detection outputs redrawn for this application. In particular, the spotting mean Average Precision on THUMOS14 is significantly improved from 59.7% to 65.6% while skipping 23% of video.

#### Moto: Enhancing Embedding with Multiple Joint Factors for Chinese Text Classification

Xunzhu Tang, Rujie Zhu, Tiezhu Sun
Track 1: Artificial Intelligence, Machine Learning for Pattern Analysis
Fri 15 Jan 2021 at 15:00 in session PS T1.14

Auto-TLDR; Moto: Enhancing Embedding with Multiple J\textbf{o}int Fac\textBF{to}rs

Recently, language representation techniques have achieved great performances in text classification. However, most existing representation models are specifically designed for English materials, which may fail in Chinese because of the huge difference between these two languages. Actually, few existing methods for Chinese text classification process texts at a single level. However, as a special kind of hieroglyphics, radicals of Chinese characters are good semantic carriers. In addition, Pinyin codes carry the semantic of tones, and Wubi reflects the stroke structure information, \textit{etc}. Unfortunately, previous researches neglected to find an effective way to distill the useful parts of these four factors and to fuse them. In our works, we propose a novel model called Moto: Enhancing Embedding with \textbf{M}ultiple J\textbf{o}int Fac\textbf{to}rs. Specifically, we design an attention mechanism to distill the useful parts by fusing the four-level information above more effectively. We conduct extensive experiments on four popular tasks. The empirical results show that our Moto achieves SOTA 0.8316 ($F_1$-score, 2.11\% improvement) on Chinese news titles, 96.38 (1.24\% improvement) on Fudan Corpus and 0.9633 (3.26\% improvement) on THUCNews.

#### 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.

#### Learning Neural Textual Representations for Citation Recommendation

Thanh Binh Kieu, Inigo Jauregi Unanue, Son Bao Pham, Xuan-Hieu Phan, M. Piccardi
Track 4: Document and Media Analysis
Wed 13 Jan 2021 at 12:00 in session PS T4.2

Auto-TLDR; Sentence-BERT cascaded with Siamese and triplet networks for citation recommendation

With the rapid growth of the scientific literature, manually selecting appropriate citations for a paper is becoming increasingly challenging and time-consuming. While several approaches for automated citation recommendation have been proposed in the recent years, effective document representations for citation recommendation are still elusive to a large extent. For this reason, in this paper we propose a novel approach to citation recommendation which leverages a deep sequential representation of the documents (Sentence-BERT) cascaded with Siamese and triplet networks in a submodular scoring function. To the best of our knowledge, this is the first approach to combine deep representations and submodular selection for a task of citation recommendation. Experiments have been carried out using a popular benchmark dataset -- the ACL Anthology Network corpus -- and evaluated against baselines and a state-of-the-art approach using metrics such as the MRR and F1@k score. The results show that the proposed approach has been able to outperform all the compared approaches in every measured metric.

#### Automatic Student Network Search for Knowledge Distillation

Zhexi Zhang, Wei Zhu, Junchi Yan, Peng Gao, Guotong Xie
Track 1: Artificial Intelligence, Machine Learning for Pattern Analysis
Thu 14 Jan 2021 at 12:00 in session PS T1.10

Auto-TLDR; NAS-KD: Knowledge Distillation for BERT

Pre-trained language models (PLMs), such as BERT, have achieved outstanding performance on multiple natural language processing (NLP) tasks. However, such pre-trained models usually contain a huge number of parameters and are computationally expensive. The high resource demand hinders their application on resource-restricted devices like mobile phones. Knowledge distillation (KD) is an effective compression approach, aiming at encouraging a light-weight student network to imitate the teacher network, and accordingly latent knowledge is transferred from the teacher to student. However, the great majority of student networks in previous KD methods are manually designed, normally a subnetwork of the teacher network. Transformer is generally utilized as the student for compressing BERT but still contains masses of parameters. Motivated by this, we propose a novel approach named NAS-KD, which automatically generates an optimal student network using neural architecture search (NAS) to enhance the distillation for BERT. Experiment on 7 classification tasks in NLP domain demonstrates that NAS-KD can substantially reduce the size of BERT without much performance sacrifice.

#### A Multi-Head Self-Relation Network for Scene Text Recognition

Zhou Junwei, Hongchao Gao, Jiao Dai, Dongqin Liu, Jizhong Han
Track 4: Document and Media Analysis
Thu 14 Jan 2021 at 12:00 in session PS T4.3

Auto-TLDR; Multi-head Self-relation Network for Scene Text Recognition

The text embedded in scene images can be seen everywhere in our lives. However, recognizing text from natural scene images is still a challenge because of its diverse shapes and distorted patterns. Recently, advanced recognition networks generally treat scene text recognition as a sequence prediction task. Although achieving excellent performance, these recognition networks consider the feature map cells as independent individuals and update cells state without utilizing the information of their neighboring cells. And the local receptive field of traditional convolutional neural network (CNN) makes a single cell that cannot cover the whole text region in an image. Due to these issues, the existing recognition networks cannot extract the global context in a visual scene. To deal with the above problems, we propose a Multi-head Self-relation Network(MSRN) for scene text recognition in this paper. The MSRN consists of several multi-head self-relation layers, which is designed for extracting the global context of a visual scene, so that transforms a cell into a new cell that fuses the information of the related cells. Furthermore, experiments over several public datasets demonstrate that our proposed recognition network achieves superior performance on several benchmark datasets including IC03, IC13, IC15, SVT-Perspective.