Shuhui Wang

Weidong Chen, Dexiang Hong, Yuankai Qi et al.

Referring video object segmentation aims to segment the object referred by a given language expression. Existing works typically require compressed video bitstream to be decoded to RGB frames before being segmented, which increases computation and storage requirements and ultimately slows the inference down. This may hamper its application in real-world computing resource limited scenarios, such as autonomous cars and drones. To alleviate this problem, in this paper, we explore the referring object segmentation task on compressed videos, namely on the original video data flow. Besides the inherent difficulty of the video referring object segmentation task itself, obtaining discriminative representation from compressed video is also rather challenging. To address this problem, we propose a multi-attention network which consists of dual-path dual-attention module and a query-based cross-modal Transformer module. Specifically, the dual-path dual-attention module is designed to extract effective representation from compressed data in three modalities, i.e., I-frame, Motion Vector and Residual. The query-based cross-modal Transformer firstly models the correlation between linguistic and visual modalities, and then the fused multi-modality features are used to guide object queries to generate a content-aware dynamic kernel and to predict final segmentation masks. Different from previous works, we propose to learn just one kernel, which thus removes the complicated post mask-matching procedure of existing methods. Extensive promising experimental results on three challenging datasets show the effectiveness of our method compared against several state-of-the-art methods which are proposed for processing RGB data. Source code is available at: https://github.com/DexiangHong/MANet.

Xiaodan Li, Yuefeng Chen, Yao Zhu et al.

Recent studies have shown that higher accuracy on ImageNet usually leads to better robustness against different corruptions. Therefore, in this paper, instead of following the traditional research paradigm that investigates new out-of-distribution corruptions or perturbations deep models may encounter, we conduct model debugging in in-distribution data to explore which object attributes a model may be sensitive to. To achieve this goal, we create a toolkit for object editing with controls of backgrounds, sizes, positions, and directions, and create a rigorous benchmark named ImageNet-E(diting) for evaluating the image classifier robustness in terms of object attributes. With our ImageNet-E, we evaluate the performance of current deep learning models, including both convolutional neural networks and vision transformers. We find that most models are quite sensitive to attribute changes. A small change in the background can lead to an average of 9.23\% drop on top-1 accuracy. We also evaluate some robust models including both adversarially trained models and other robust trained models and find that some models show worse robustness against attribute changes than vanilla models. Based on these findings, we discover ways to enhance attribute robustness with preprocessing, architecture designs, and training strategies. We hope this work can provide some insights to the community and open up a new avenue for research in robust computer vision. The code and dataset are available at https://github.com/alibaba/easyrobust.

Zhenhuan Liu, Jincan Deng, Liang Li et al. · nvidia, pku

Conditional image generation is an active research topic including text2image and image translation. Recently image manipulation with linguistic instruction brings new challenges of multimodal conditional generation. However, traditional conditional image generation models mainly focus on generating high-quality and visually realistic images, and lack resolving the partial consistency between image and instruction. To address this issue, we propose an Increment Reasoning Generative Adversarial Network (IR-GAN), which aims to reason the consistency between visual increment in images and semantic increment in instructions. First, we introduce the word-level and instruction-level instruction encoders to learn user's intention from history-correlated instructions as semantic increment. Second, we embed the representation of semantic increment into that of source image for generating target image, where source image plays the role of referring auxiliary. Finally, we propose a reasoning discriminator to measure the consistency between visual increment and semantic increment, which purifies user's intention and guarantees the good logic of generated target image. Extensive experiments and visualization conducted on two datasets show the effectiveness of IR-GAN.

Zhenhuan Liu, Liang Li, Huajie Jiang et al. · nvidia

In recent years, creative content generations like style transfer and neural photo editing have attracted more and more attention. Among these, cartoonization of real-world scenes has promising applications in entertainment and industry. Different from image translations focusing on improving the style effect of generated images, video cartoonization has additional requirements on the temporal consistency. In this paper, we propose a spatially-adaptive semantic alignment framework with perceptual motion consistency for coherent video cartoonization in an unsupervised manner. The semantic alignment module is designed to restore deformation of semantic structure caused by spatial information lost in the encoder-decoder architecture. Furthermore, we devise the spatio-temporal correlative map as a style-independent, global-aware regularization on the perceptual motion consistency. Deriving from similarity measurement of high-level features in photo and cartoon frames, it captures global semantic information beyond raw pixel-value in optical flow. Besides, the similarity measurement disentangles temporal relationships from domain-specific style properties, which helps regularize the temporal consistency without hurting style effects of cartoon images. Qualitative and quantitative experiments demonstrate our method is able to generate highly stylistic and temporal consistent cartoon videos.

Chen Zhang, Guorong Li, Yuankai Qi et al.

Weakly supervised video anomaly detection aims to identify abnormal events in videos using only video-level labels. Recently, two-stage self-training methods have achieved significant improvements by self-generating pseudo labels and self-refining anomaly scores with these labels. As the pseudo labels play a crucial role, we propose an enhancement framework by exploiting completeness and uncertainty properties for effective self-training. Specifically, we first design a multi-head classification module (each head serves as a classifier) with a diversity loss to maximize the distribution differences of predicted pseudo labels across heads. This encourages the generated pseudo labels to cover as many abnormal events as possible. We then devise an iterative uncertainty pseudo label refinement strategy, which improves not only the initial pseudo labels but also the updated ones obtained by the desired classifier in the second stage. Extensive experimental results demonstrate the proposed method performs favorably against state-of-the-art approaches on the UCF-Crime, TAD, and XD-Violence benchmark datasets.

Xuejing Liu, Liang Li, Shuhui Wang et al.

Weakly supervised Referring Expression Grounding (REG) aims to ground a particular target in an image described by a language expression while lacking the correspondence between target and expression. Two main problems exist in weakly supervised REG. First, the lack of region-level annotations introduces ambiguities between proposals and queries. Second, most previous weakly supervised REG methods ignore the discriminative location and context of the referent, causing difficulties in distinguishing the target from other same-category objects. To address the above challenges, we design an entity-enhanced adaptive reconstruction network (EARN). Specifically, EARN includes three modules: entity enhancement, adaptive grounding, and collaborative reconstruction. In entity enhancement, we calculate semantic similarity as supervision to select the candidate proposals. Adaptive grounding calculates the ranking score of candidate proposals upon subject, location and context with hierarchical attention. Collaborative reconstruction measures the ranking result from three perspectives: adaptive reconstruction, language reconstruction and attribute classification. The adaptive mechanism helps to alleviate the variance of different referring expressions. Experiments on five datasets show EARN outperforms existing state-of-the-art methods. Qualitative results demonstrate that the proposed EARN can better handle the situation where multiple objects of a particular category are situated together.

Yabing Wang, Shuhui Wang, Hao Luo et al. · stanford

Current research on cross-modal retrieval is mostly English-oriented, as the availability of a large number of English-oriented human-labeled vision-language corpora. In order to break the limit of non-English labeled data, cross-lingual cross-modal retrieval (CCR) has attracted increasing attention. Most CCR methods construct pseudo-parallel vision-language corpora via Machine Translation (MT) to achieve cross-lingual transfer. However, the translated sentences from MT are generally imperfect in describing the corresponding visual contents. Improperly assuming the pseudo-parallel data are correctly correlated will make the networks overfit to the noisy correspondence. Therefore, we propose Dual-view Curricular Optimal Transport (DCOT) to learn with noisy correspondence in CCR. In particular, we quantify the confidence of the sample pair correlation with optimal transport theory from both the cross-lingual and cross-modal views, and design dual-view curriculum learning to dynamically model the transportation costs according to the learning stage of the two views. Extensive experiments are conducted on two multilingual image-text datasets and one video-text dataset, and the results demonstrate the effectiveness and robustness of the proposed method. Besides, our proposed method also shows a good expansibility to cross-lingual image-text baselines and a decent generalization on out-of-domain data.

Junshu Sun, Shuhui Wang, Chenxue Yang et al.

Designing effective graph neural networks (GNNs) with message passing has two fundamental challenges, i.e., determining optimal message-passing pathways and designing local aggregators. Previous methods of designing optimal pathways are limited with information loss on the input features. On the other hand, existing local aggregators generally fail to extract multi-scale features and approximate diverse operators under limited parameter scales. In contrast to these methods, Euclidean convolution has been proven as an expressive aggregator, making it a perfect candidate for GNN construction. However, the challenges of generalizing Euclidean convolution to graphs arise from the irregular structure of graphs. To bridge the gap between Euclidean space and graph topology, we propose a differentiable method that applies permutations to calibrate input graphs for Euclidean convolution. The permutations constrain all nodes in a row regardless of their input order and therefore enable the flexible generalization of Euclidean convolution to graphs. Based on the graph calibration, we propose the Compressed Convolution Network (CoCN) for hierarchical graph representation learning. CoCN follows local feature-learning and global parameter-sharing mechanisms of convolution neural networks. The whole model can be trained end-to-end, with compressed convolution applied to learn individual node features and their corresponding structure features. CoCN can further borrow successful practices from Euclidean convolution, including residual connection and inception mechanism. We validate CoCN on both node-level and graph-level benchmarks. CoCN achieves superior performance over competitive GNN baselines. Codes are available at https://github.com/sunjss/CoCN.

Guanqi Ding, Xinzhe Han, Shuhui Wang et al.

Few-shot image generation aims to generate data of an unseen category based on only a few samples. Apart from basic content generation, a bunch of downstream applications hopefully benefit from this task, such as low-data detection and few-shot classification. To achieve this goal, the generated images should guarantee category retention for classification beyond the visual quality and diversity. In our preliminary work, we present an ``editing-based'' framework Attribute Group Editing (AGE) for reliable few-shot image generation, which largely improves the generation performance. Nevertheless, AGE's performance on downstream classification is not as satisfactory as expected. This paper investigates the class inconsistency problem and proposes Stable Attribute Group Editing (SAGE) for more stable class-relevant image generation. SAGE takes use of all given few-shot images and estimates a class center embedding based on the category-relevant attribute dictionary. Meanwhile, according to the projection weights on the category-relevant attribute dictionary, we can select category-irrelevant attributes from the similar seen categories. Consequently, SAGE injects the whole distribution of the novel class into StyleGAN's latent space, thus largely remains the category retention and stability of the generated images. Going one step further, we find that class inconsistency is a common problem in GAN-generated images for downstream classification. Even though the generated images look photo-realistic and requires no category-relevant editing, they are usually of limited help for downstream classification. We systematically discuss this issue from both the generative model and classification model perspectives, and propose to boost the downstream classification performance of SAGE by enhancing the pixel and frequency components.

Jiayu Xiao, Henglei Lv, Liang Li et al.

Recent text-to-image (T2I) diffusion models have achieved remarkable progress in generating high-quality images given text-prompts as input. However, these models fail to convey appropriate spatial composition specified by a layout instruction. In this work, we probe into zero-shot grounded T2I generation with diffusion models, that is, generating images corresponding to the input layout information without training auxiliary modules or finetuning diffusion models. We propose a Region and Boundary (R&B) aware cross-attention guidance approach that gradually modulates the attention maps of diffusion model during generative process, and assists the model to synthesize images (1) with high fidelity, (2) highly compatible with textual input, and (3) interpreting layout instructions accurately. Specifically, we leverage the discrete sampling to bridge the gap between consecutive attention maps and discrete layout constraints, and design a region-aware loss to refine the generative layout during diffusion process. We further propose a boundary-aware loss to strengthen object discriminability within the corresponding regions. Experimental results show that our method outperforms existing state-of-the-art zero-shot grounded T2I generation methods by a large margin both qualitatively and quantitatively on several benchmarks.

Yan Zhu, Junbao Zhuo, Bin Ma et al.

Zero-shot video recognition (ZSVR) is a task that aims to recognize video categories that have not been seen during the model training process. Recently, vision-language models (VLMs) pre-trained on large-scale image-text pairs have demonstrated impressive transferability for ZSVR. To make VLMs applicable to the video domain, existing methods often use an additional temporal learning module after the image-level encoder to learn the temporal relationships among video frames. Unfortunately, for video from unseen categories, we observe an abnormal phenomenon where the model that uses spatial-temporal feature performs much worse than the model that removes temporal learning module and uses only spatial feature. We conjecture that improper temporal modeling on video disrupts the spatial feature of the video. To verify our hypothesis, we propose Feature Factorization to retain the orthogonal temporal feature of the video and use interpolation to construct refined spatial-temporal feature. The model using appropriately refined spatial-temporal feature performs better than the one using only spatial feature, which verifies the effectiveness of the orthogonal temporal feature for the ZSVR task. Therefore, an Orthogonal Temporal Interpolation module is designed to learn a better refined spatial-temporal video feature during training. Additionally, a Matching Loss is introduced to improve the quality of the orthogonal temporal feature. We propose a model called OTI for ZSVR by employing orthogonal temporal interpolation and the matching loss based on VLMs. The ZSVR accuracies on popular video datasets (i.e., Kinetics-600, UCF101 and HMDB51) show that OTI outperforms the previous state-of-the-art method by a clear margin.

Guanqi Ding, Xinzhe Han, Shuhui Wang et al.

Few-shot image generation is a challenging task even using the state-of-the-art Generative Adversarial Networks (GANs). Due to the unstable GAN training process and the limited training data, the generated images are often of low quality and low diversity. In this work, we propose a new editing-based method, i.e., Attribute Group Editing (AGE), for few-shot image generation. The basic assumption is that any image is a collection of attributes and the editing direction for a specific attribute is shared across all categories. AGE examines the internal representation learned in GANs and identifies semantically meaningful directions. Specifically, the class embedding, i.e., the mean vector of the latent codes from a specific category, is used to represent the category-relevant attributes, and the category-irrelevant attributes are learned globally by Sparse Dictionary Learning on the difference between the sample embedding and the class embedding. Given a GAN well trained on seen categories, diverse images of unseen categories can be synthesized through editing category-irrelevant attributes while keeping category-relevant attributes unchanged. Without re-training the GAN, AGE is capable of not only producing more realistic and diverse images for downstream visual applications with limited data but achieving controllable image editing with interpretable category-irrelevant directions.

Jingru Gan, Xinzhe Han, Shuhui Wang et al.

Given an image and an associated textual question, the purpose of Knowledge-Based Visual Question Answering (KB-VQA) is to provide a correct answer to the question with the aid of external knowledge bases. Prior KB-VQA models are usually formulated as a retriever-classifier framework, where a pre-trained retriever extracts textual or visual information from knowledge graphs and then makes a prediction among the candidates. Despite promising progress, there are two drawbacks with existing models. Firstly, modeling question-answering as multi-class classification limits the answer space to a preset corpus and lacks the ability of flexible reasoning. Secondly, the classifier merely consider "what is the answer" without "how to get the answer", which cannot ground the answer to explicit reasoning paths. In this paper, we confront the challenge of \emph{explainable open-set} KB-VQA, where the system is required to answer questions with entities at wild and retain an explainable reasoning path. To resolve the aforementioned issues, we propose a new retriever-ranker paradigm of KB-VQA, Graph pATH rankER (GATHER for brevity). Specifically, it contains graph constructing, pruning, and path-level ranking, which not only retrieves accurate answers but also provides inference paths that explain the reasoning process. To comprehensively evaluate our model, we reformulate the benchmark dataset OK-VQA with manually corrected entity-level annotations and release it as ConceptVQA. Extensive experiments on real-world questions demonstrate that our framework is not only able to perform open-set question answering across the whole knowledge base but provide explicit reasoning path.

Zhaobo Qi, Shuhui Wang, Weigang Zhang et al.

Video activity anticipation aims to predict what will happen in the future, embracing a broad application prospect ranging from robot vision and autonomous driving. Despite the recent progress, the data uncertainty issue, reflected as the content evolution process and dynamic correlation in event labels, has been somehow ignored. This reduces the model generalization ability and deep understanding on video content, leading to serious error accumulation and degraded performance. In this paper, we address the uncertainty learning problem and propose an uncertainty-boosted robust video activity anticipation framework, which generates uncertainty values to indicate the credibility of the anticipation results. The uncertainty value is used to derive a temperature parameter in the softmax function to modulate the predicted target activity distribution. To guarantee the distribution adjustment, we construct a reasonable target activity label representation by incorporating the activity evolution from the temporal class correlation and the semantic relationship. Moreover, we quantify the uncertainty into relative values by comparing the uncertainty among sample pairs and their temporal-lengths. This relative strategy provides a more accessible way in uncertainty modeling than quantifying the absolute uncertainty values on the whole dataset. Experiments on multiple backbones and benchmarks show our framework achieves promising performance and better robustness/interpretability. Source codes are available at https://github.com/qzhb/UbRV2A.

Zhaobo Qi, Yibo Yuan, Xiaowen Ruan et al.

Temporal Sentence Grounding in Video (TSGV) is troubled by dataset bias issue, which is caused by the uneven temporal distribution of the target moments for samples with similar semantic components in input videos or query texts. Existing methods resort to utilizing prior knowledge about bias to artificially break this uneven distribution, which only removes a limited amount of significant language biases. In this work, we propose the bias-conflict sample synthesis and adversarial removal debias strategy (BSSARD), which dynamically generates bias-conflict samples by explicitly leveraging potentially spurious correlations between single-modality features and the temporal position of the target moments. Through adversarial training, its bias generators continuously introduce biases and generate bias-conflict samples to deceive its grounding model. Meanwhile, the grounding model continuously eliminates the introduced biases, which requires it to model multi-modality alignment information. BSSARD will cover most kinds of coupling relationships and disrupt language and visual biases simultaneously. Extensive experiments on Charades-CD and ActivityNet-CD demonstrate the promising debiasing capability of BSSARD. Source codes are available at https://github.com/qzhb/BSSARD.

Yufan Zhou, Zhaobo Qi, Lingshuai Lin et al.

In this paper, we address the challenge of procedure planning in instructional videos, aiming to generate coherent and task-aligned action sequences from start and end visual observations. Previous work has mainly relied on text-level supervision to bridge the gap between observed states and unobserved actions, but it struggles with capturing intricate temporal relationships among actions. Building on these efforts, we propose the Masked Temporal Interpolation Diffusion (MTID) model that introduces a latent space temporal interpolation module within the diffusion model. This module leverages a learnable interpolation matrix to generate intermediate latent features, thereby augmenting visual supervision with richer mid-state details. By integrating this enriched supervision into the model, we enable end-to-end training tailored to task-specific requirements, significantly enhancing the model's capacity to predict temporally coherent action sequences. Additionally, we introduce an action-aware mask projection mechanism to restrict the action generation space, combined with a task-adaptive masked proximity loss to prioritize more accurate reasoning results close to the given start and end states over those in intermediate steps. Simultaneously, it filters out task-irrelevant action predictions, leading to contextually aware action sequences. Experimental results across three widely used benchmark datasets demonstrate that our MTID achieves promising action planning performance on most metrics. The code is available at https://github.com/WiserZhou/MTID.

Churan Zhi, Junbao Zhuo, Shuhui Wang

In this paper, we address unsupervised domain adaptation under noisy environments, which is more challenging and practical than traditional domain adaptation. In this scenario, the model is prone to overfitting noisy labels, resulting in a more pronounced domain shift and a notable decline in the overall model performance. Previous methods employed prototype methods for domain adaptation on robust feature spaces. However, these approaches struggle to effectively classify classes with similar features under noisy environments. To address this issue, we propose a new method to detect and correct confusing class pair. We first divide classes into easy and hard classes based on the small loss criterion. We then leverage the top-2 predictions for each sample after aligning the source and target domain to find the confusing pair in the hard classes. We apply label correction to the noisy samples within the confusing pair. With the proposed label correction method, we can train our model with more accurate labels. Extensive experiments confirm the effectiveness of our method and demonstrate its favorable performance compared with existing state-of-the-art methods. Our codes are publicly available at https://github.com/Hehxcf/CPC/.

Xu Yan, Zhengcong Fei, Zekang Li et al.

Current state-of-the-art approaches for image captioning typically adopt an autoregressive manner, i.e., generating descriptions word by word, which suffers from slow decoding issue and becomes a bottleneck in real-time applications. Non-autoregressive image captioning with continuous iterative refinement, which eliminates the sequential dependence in a sentence generation, can achieve comparable performance to the autoregressive counterparts with a considerable acceleration. Nevertheless, based on a well-designed experiment, we empirically proved that iteration times can be effectively reduced when providing sufficient prior knowledge for the language decoder. Towards that end, we propose a novel two-stage framework, referred to as Semi-Autoregressive Image Captioning (SAIC), to make a better trade-off between performance and speed. The proposed SAIC model maintains autoregressive property in global but relieves it in local. Specifically, SAIC model first jumpily generates an intermittent sequence in an autoregressive manner, that is, it predicts the first word in every word group in order. Then, with the help of the partially deterministic prior information and image features, SAIC model non-autoregressively fills all the skipped words with one iteration. Experimental results on the MS COCO benchmark demonstrate that our SAIC model outperforms the preceding non-autoregressive image captioning models while obtaining a competitive inference speedup. Code is available at https://github.com/feizc/SAIC.

Shuhao Cui, Shuhui Wang, Junbao Zhuo et al.

Due to the domain discrepancy in visual domain adaptation, the performance of source model degrades when bumping into the high data density near decision boundary in target domain. A common solution is to minimize the Shannon Entropy to push the decision boundary away from the high density area. However, entropy minimization also leads to severe reduction of prediction diversity, and unfortunately brings harm to the domain adaptation. In this paper, we investigate the prediction discriminability and diversity by studying the structure of the classification output matrix of a randomly selected data batch. We find by theoretical analysis that the prediction discriminability and diversity could be separately measured by the Frobenius-norm and rank of the batch output matrix. The nuclear-norm is an upperbound of the former, and a convex approximation of the latter. Accordingly, we propose Batch Nuclear-norm Maximization and Minimization, which performs nuclear-norm maximization on the target output matrix to enhance the target prediction ability, and nuclear-norm minimization on the source batch output matrix to increase applicability of the source domain knowledge. We further approximate the nuclear-norm by L_{1,2}-norm, and design multi-batch optimization for stable solution on large number of categories. The fast approximation method achieves O(n^2) computational complexity and better convergence property. Experiments show that our method could boost the adaptation accuracy and robustness under three typical domain adaptation scenarios. The code is available at https://github.com/cuishuhao/BNM.

Shuhao Cui, Xuan Jin, Shuhui Wang et al.

In visual domain adaptation (DA), separating the domain-specific characteristics from the domain-invariant representations is an ill-posed problem. Existing methods apply different kinds of priors or directly minimize the domain discrepancy to address this problem, which lack flexibility in handling real-world situations. Another research pipeline expresses the domain-specific information as a gradual transferring process, which tends to be suboptimal in accurately removing the domain-specific properties. In this paper, we address the modeling of domain-invariant and domain-specific information from the heuristic search perspective. We identify the characteristics in the existing representations that lead to larger domain discrepancy as the heuristic representations. With the guidance of heuristic representations, we formulate a principled framework of Heuristic Domain Adaptation (HDA) with well-founded theoretical guarantees. To perform HDA, the cosine similarity scores and independence measurements between domain-invariant and domain-specific representations are cast into the constraints at the initial and final states during the learning procedure. Similar to the final condition of heuristic search, we further derive a constraint enforcing the final range of heuristic network output to be small. Accordingly, we propose Heuristic Domain Adaptation Network (HDAN), which explicitly learns the domain-invariant and domain-specific representations with the above mentioned constraints. Extensive experiments show that HDAN has exceeded state-of-the-art on unsupervised DA, multi-source DA and semi-supervised DA. The code is available at https://github.com/cuishuhao/HDA.

Shuhao Cui, Shuhui Wang, Junbao Zhuo et al.

In unsupervised domain adaptation, rich domain-specific characteristics bring great challenge to learn domain-invariant representations. However, domain discrepancy is considered to be directly minimized in existing solutions, which is difficult to achieve in practice. Some methods alleviate the difficulty by explicitly modeling domain-invariant and domain-specific parts in the representations, but the adverse influence of the explicit construction lies in the residual domain-specific characteristics in the constructed domain-invariant representations. In this paper, we equip adversarial domain adaptation with Gradually Vanishing Bridge (GVB) mechanism on both generator and discriminator. On the generator, GVB could not only reduce the overall transfer difficulty, but also reduce the influence of the residual domain-specific characteristics in domain-invariant representations. On the discriminator, GVB contributes to enhance the discriminating ability, and balance the adversarial training process. Experiments on three challenging datasets show that our GVB methods outperform strong competitors, and cooperate well with other adversarial methods. The code is available at https://github.com/cuishuhao/GVB.

Shuhao Cui, Shuhui Wang, Junbao Zhuo et al.

The learning of the deep networks largely relies on the data with human-annotated labels. In some label insufficient situations, the performance degrades on the decision boundary with high data density. A common solution is to directly minimize the Shannon Entropy, but the side effect caused by entropy minimization, i.e., reduction of the prediction diversity, is mostly ignored. To address this issue, we reinvestigate the structure of classification output matrix of a randomly selected data batch. We find by theoretical analysis that the prediction discriminability and diversity could be separately measured by the Frobenius-norm and rank of the batch output matrix. Besides, the nuclear-norm is an upperbound of the Frobenius-norm, and a convex approximation of the matrix rank. Accordingly, to improve both discriminability and diversity, we propose Batch Nuclear-norm Maximization (BNM) on the output matrix. BNM could boost the learning under typical label insufficient learning scenarios, such as semi-supervised learning, domain adaptation and open domain recognition. On these tasks, extensive experimental results show that BNM outperforms competitors and works well with existing well-known methods. The code is available at https://github.com/cuishuhao/BNM.

Ting Yu, Xiaojun Lin, Shuhui Wang et al.

Three-Dimensional (3D) dense captioning is an emerging vision-language bridging task that aims to generate multiple detailed and accurate descriptions for 3D scenes. It presents significant potential and challenges due to its closer representation of the real world compared to 2D visual captioning, as well as complexities in data collection and processing of 3D point cloud sources. Despite the popularity and success of existing methods, there is a lack of comprehensive surveys summarizing the advancements in this field, which hinders its progress. In this paper, we provide a comprehensive review of 3D dense captioning, covering task definition, architecture classification, dataset analysis, evaluation metrics, and in-depth prosperity discussions. Based on a synthesis of previous literature, we refine a standard pipeline that serves as a common paradigm for existing methods. We also introduce a clear taxonomy of existing models, summarize technologies involved in different modules, and conduct detailed experiment analysis. Instead of a chronological order introduction, we categorize the methods into different classes to facilitate exploration and analysis of the differences and connections among existing techniques. We also provide a reading guideline to assist readers with different backgrounds and purposes in reading efficiently. Furthermore, we propose a series of promising future directions for 3D dense captioning by identifying challenges and aligning them with the development of related tasks, offering valuable insights and inspiring future research in this field. Our aim is to provide a comprehensive understanding of 3D dense captioning, foster further investigations, and contribute to the development of novel applications in multimedia and related domains.

Junshu Sun, Chenxue Yang, Xiangyang Ji et al.

Message passing plays a vital role in graph neural networks (GNNs) for effective feature learning. However, the over-reliance on input topology diminishes the efficacy of message passing and restricts the ability of GNNs. Despite efforts to mitigate the reliance, existing study encounters message-passing bottlenecks or high computational expense problems, which invokes the demands for flexible message passing with low complexity. In this paper, we propose a novel dynamic message-passing mechanism for GNNs. It projects graph nodes and learnable pseudo nodes into a common space with measurable spatial relations between them. With nodes moving in the space, their evolving relations facilitate flexible pathway construction for a dynamic message-passing process. Associating pseudo nodes to input graphs with their measured relations, graph nodes can communicate with each other intermediately through pseudo nodes under linear complexity. We further develop a GNN model named $\mathtt{\mathbf{N^2}}$ based on our dynamic message-passing mechanism. $\mathtt{\mathbf{N^2}}$ employs a single recurrent layer to recursively generate the displacements of nodes and construct optimal dynamic pathways. Evaluation on eighteen benchmarks demonstrates the superior performance of $\mathtt{\mathbf{N^2}}$ over popular GNNs. $\mathtt{\mathbf{N^2}}$ successfully scales to large-scale benchmarks and requires significantly fewer parameters for graph classification with the shared recurrent layer.

Ting Yu, Kunhao Fu, Shuhui Wang et al.

Video Question Answering (VideoQA) represents a crucial intersection between video understanding and language processing, requiring both discriminative unimodal comprehension and sophisticated cross-modal interaction for accurate inference. Despite advancements in multi-modal pre-trained models and video-language foundation models, these systems often struggle with domain-specific VideoQA due to their generalized pre-training objectives. Addressing this gap necessitates bridging the divide between broad cross-modal knowledge and the specific inference demands of VideoQA tasks. To this end, we introduce HeurVidQA, a framework that leverages domain-specific entity-action heuristics to refine pre-trained video-language foundation models. Our approach treats these models as implicit knowledge engines, employing domain-specific entity-action prompters to direct the model's focus toward precise cues that enhance reasoning. By delivering fine-grained heuristics, we improve the model's ability to identify and interpret key entities and actions, thereby enhancing its reasoning capabilities. Extensive evaluations across multiple VideoQA datasets demonstrate that our method significantly outperforms existing models, underscoring the importance of integrating domain-specific knowledge into video-language models for more accurate and context-aware VideoQA.

Zhengqi Pei, Qingming Huang, Shuhui Wang · utoronto

The ever-increasing scale of modern neural networks has brought unprecedented performance alongside daunting challenges in efficiency and interpretability. This paper addresses the core question of how to build large neural systems that learn efficient, modular, and interpretable representations. We propose Neuronal Group Communication (NGC), a theory-driven framework that reimagines a neural network as a dynamical system of interacting neuronal groups rather than a monolithic collection of neural weights. Instead of treating each weight as an independent trainable parameter, NGC treats weights as transient interactions between embedding-like neuronal states, with neural computation unfolding through iterative communication among groups of neurons. This low-rank, modular representation yields compact models: groups of neurons exchange low-dimensional signals, enabling intra-group specialization and inter-group information sharing while dramatically reducing redundant parameters. By drawing on dynamical systems theory, we introduce a neuronal stability metric (analogous to Lyapunov stability) that quantifies the contraction of neuron activations toward stable patterns during sequence processing. Using this metric, we reveal that emergent reasoning capabilities correspond to an external driving force or ``potential'', which nudges the neural dynamics away from trivial trajectories while preserving stability. Empirically, we instantiate NGC in large language models (LLMs) and demonstrate improved performance on complex reasoning benchmarks under moderate compression. NGC consistently outperforms standard low-rank approximations and cross-layer basis-sharing methods at comparable compression rates. We conclude by discussing the broader implications of NGC, including how structured neuronal group dynamics might relate to generalization in high-dimensional learning systems.

Shuo Cai, Xinzhe Han, Shuhui Wang

Goal-oriented visual dialogue involves multi-round interaction between artificial agents, which has been of remarkable attention due to its wide applications. Given a visual scene, this task occurs when a Questioner asks an action-oriented question and an Answerer responds with the intent of letting the Questioner know the correct action to take. The quality of questions affects the accuracy and efficiency of the target search progress. However, existing methods lack a clear strategy to guide the generation of questions, resulting in the randomness in the search process and inconvergent results. We propose a Tree-Structured Strategy with Answer Distribution Estimator (TSADE) which guides the question generation by excluding half of the current candidate objects in each round. The above process is implemented by maximizing a binary reward inspired by the ``divide-and-conquer'' paradigm. We further design a candidate-minimization reward which encourages the model to narrow down the scope of candidate objects toward the end of the dialogue. We experimentally demonstrate that our method can enable the agents to achieve high task-oriented accuracy with fewer repeating questions and rounds compared to traditional ergodic question generation approaches. Qualitative results further show that TSADE facilitates agents to generate higher-quality questions.

Xinzhe Han, Shuhui Wang, Chi Su et al.

Neural networks often make predictions relying on the spurious correlations from the datasets rather than the intrinsic properties of the task of interest, facing sharp degradation on out-of-distribution (OOD) test data. Existing de-bias learning frameworks try to capture specific dataset bias by annotations but they fail to handle complicated OOD scenarios. Others implicitly identify the dataset bias by special design low capability biased models or losses, but they degrade when the training and testing data are from the same distribution. In this paper, we propose a General Greedy De-bias learning framework (GGD), which greedily trains the biased models and the base model. The base model is encouraged to focus on examples that are hard to solve with biased models, thus remaining robust against spurious correlations in the test stage. GGD largely improves models' OOD generalization ability on various tasks, but sometimes over-estimates the bias level and degrades on the in-distribution test. We further re-analyze the ensemble process of GGD and introduce the Curriculum Regularization inspired by curriculum learning, which achieves a good trade-off between in-distribution and out-of-distribution performance. Extensive experiments on image classification, adversarial question answering, and visual question answering demonstrate the effectiveness of our method. GGD can learn a more robust base model under the settings of both task-specific biased models with prior knowledge and self-ensemble biased model without prior knowledge.

Hanhua Ye, Guorong Li, Yuankai Qi et al.

Video captioning aims to generate natural language descriptions according to the content, where representation learning plays a crucial role. Existing methods are mainly developed within the supervised learning framework via word-by-word comparison of the generated caption against the ground-truth text without fully exploiting linguistic semantics. In this work, we propose a hierarchical modular network to bridge video representations and linguistic semantics from three levels before generating captions. In particular, the hierarchy is composed of: (I) Entity level, which highlights objects that are most likely to be mentioned in captions. (II) Predicate level, which learns the actions conditioned on highlighted objects and is supervised by the predicate in captions. (III) Sentence level, which learns the global semantic representation and is supervised by the whole caption. Each level is implemented by one module. Extensive experimental results show that the proposed method performs favorably against the state-of-the-art models on the two widely-used benchmarks: MSVD 104.0% and MSR-VTT 51.5% in CIDEr score.

Zhaobo Qi, Shuhui Wang, Chi Su et al.

Future activity anticipation is a challenging problem in egocentric vision. As a standard future activity anticipation paradigm, recursive sequence prediction suffers from the accumulation of errors. To address this problem, we propose a simple and effective Self-Regulated Learning framework, which aims to regulate the intermediate representation consecutively to produce representation that (a) emphasizes the novel information in the frame of the current time-stamp in contrast to previously observed content, and (b) reflects its correlation with previously observed frames. The former is achieved by minimizing a contrastive loss, and the latter can be achieved by a dynamic reweighing mechanism to attend to informative frames in the observed content with a similarity comparison between feature of the current frame and observed frames. The learned final video representation can be further enhanced by multi-task learning which performs joint feature learning on the target activity labels and the automatically detected action and object class tokens. SRL sharply outperforms existing state-of-the-art in most cases on two egocentric video datasets and two third-person video datasets. Its effectiveness is also verified by the experimental fact that the action and object concepts that support the activity semantics can be accurately identified.

Xu Yan, Zhengcong Fei, Shuhui Wang et al.

Dense video captioning (DVC) aims to generate multi-sentence descriptions to elucidate the multiple events in the video, which is challenging and demands visual consistency, discoursal coherence, and linguistic diversity. Existing methods mainly generate captions from individual video segments, lacking adaptation to the global visual context and progressive alignment between the fast-evolved visual content and textual descriptions, which results in redundant and spliced descriptions. In this paper, we introduce the concept of information flow to model the progressive information changing across video sequence and captions. By designing a Cross-modal Information Flow Alignment mechanism, the visual and textual information flows are captured and aligned, which endows the captioning process with richer context and dynamics on event/topic evolution. Based on the Cross-modal Information Flow Alignment module, we further put forward DVCFlow framework, which consists of a Global-local Visual Encoder to capture both global features and local features for each video segment, and a pre-trained Caption Generator to produce captions. Extensive experiments on the popular ActivityNet Captions and YouCookII datasets demonstrate that our method significantly outperforms competitive baselines, and generates more human-like text according to subject and objective tests.

Xinzhe Han, Shuhui Wang, Chi Su et al.

Language bias is a critical issue in Visual Question Answering (VQA), where models often exploit dataset biases for the final decision without considering the image information. As a result, they suffer from performance drop on out-of-distribution data and inadequate visual explanation. Based on experimental analysis for existing robust VQA methods, we stress the language bias in VQA that comes from two aspects, i.e., distribution bias and shortcut bias. We further propose a new de-bias framework, Greedy Gradient Ensemble (GGE), which combines multiple biased models for unbiased base model learning. With the greedy strategy, GGE forces the biased models to over-fit the biased data distribution in priority, thus makes the base model pay more attention to examples that are hard to solve by biased models. The experiments demonstrate that our method makes better use of visual information and achieves state-of-the-art performance on diagnosing dataset VQA-CP without using extra annotations.

Jinghao Zhang, Yanqiao Zhu, Qiang Liu et al.

Multimedia content is of predominance in the modern Web era. Investigating how users interact with multimodal items is a continuing concern within the rapid development of recommender systems. The majority of previous work focuses on modeling user-item interactions with multimodal features included as side information. However, this scheme is not well-designed for multimedia recommendation. Specifically, only collaborative item-item relationships are implicitly modeled through high-order item-user-item relations. Considering that items are associated with rich contents in multiple modalities, we argue that the latent semantic item-item structures underlying these multimodal contents could be beneficial for learning better item representations and further boosting recommendation. To this end, we propose a LATent sTructure mining method for multImodal reCommEndation, which we term LATTICE for brevity. To be specific, in the proposed LATTICE model, we devise a novel modality-aware structure learning layer, which learns item-item structures for each modality and aggregates multiple modalities to obtain latent item graphs. Based on the learned latent graphs, we perform graph convolutions to explicitly inject high-order item affinities into item representations. These enriched item representations can then be plugged into existing collaborative filtering methods to make more accurate recommendations. Extensive experiments on three real-world datasets demonstrate the superiority of our method over state-of-the-art multimedia recommendation methods and validate the efficacy of mining latent item-item relationships from multimodal features.

Xiaodan Li, Jinfeng Li, Yuefeng Chen et al.

We study the query-based attack against image retrieval to evaluate its robustness against adversarial examples under the black-box setting, where the adversary only has query access to the top-k ranked unlabeled images from the database. Compared with query attacks in image classification, which produce adversaries according to the returned labels or confidence score, the challenge becomes even more prominent due to the difficulty in quantifying the attack effectiveness on the partial retrieved list. In this paper, we make the first attempt in Query-based Attack against Image Retrieval (QAIR), to completely subvert the top-k retrieval results. Specifically, a new relevance-based loss is designed to quantify the attack effects by measuring the set similarity on the top-k retrieval results before and after attacks and guide the gradient optimization. To further boost the attack efficiency, a recursive model stealing method is proposed to acquire transferable priors on the target model and generate the prior-guided gradients. Comprehensive experiments show that the proposed attack achieves a high attack success rate with few queries against the image retrieval systems under the black-box setting. The attack evaluations on the real-world visual search engine show that it successfully deceives a commercial system such as Bing Visual Search with 98% attack success rate by only 33 queries on average.

Xiaofeng Mao, Yuefeng Chen, Shuhui Wang et al.

Adversarial attack is a technique for deceiving Machine Learning (ML) models, which provides a way to evaluate the adversarial robustness. In practice, attack algorithms are artificially selected and tuned by human experts to break a ML system. However, manual selection of attackers tends to be sub-optimal, leading to a mistakenly assessment of model security. In this paper, a new procedure called Composite Adversarial Attack (CAA) is proposed for automatically searching the best combination of attack algorithms and their hyper-parameters from a candidate pool of \textbf{32 base attackers}. We design a search space where attack policy is represented as an attacking sequence, i.e., the output of the previous attacker is used as the initialization input for successors. Multi-objective NSGA-II genetic algorithm is adopted for finding the strongest attack policy with minimum complexity. The experimental result shows CAA beats 10 top attackers on 11 diverse defenses with less elapsed time (\textbf{6 $\times$ faster than AutoAttack}), and achieves the new state-of-the-art on $l_{\infty}$, $l_{2}$ and unrestricted adversarial attacks.

Jianfeng He, Xuchao Zhang, Shuo Lei et al.

Semantic editing on segmentation map has been proposed as an intermediate interface for image generation, because it provides flexible and strong assistance in various image generation tasks. This paper aims to improve quality of edited segmentation map conditioned on semantic inputs. Even though recent studies apply global and local adversarial losses extensively to generate images for higher image quality, we find that they suffer from the misalignment of the boundary area in the mask area. To address this, we propose MExGAN for semantic editing on segmentation map, which uses a novel Multi-Expansion (MEx) loss implemented by adversarial losses on MEx areas. Each MEx area has the mask area of the generation as the majority and the boundary of original context as the minority. To boost convenience and stability of MEx loss, we further propose an Approximated MEx (A-MEx) loss. Besides, in contrast to previous model that builds training data for semantic editing on segmentation map with part of the whole image, which leads to model performance degradation, MExGAN applies the whole image to build the training data. Extensive experiments on semantic editing on segmentation map and natural image inpainting show competitive results on four datasets.

Jun Wei, Shuhui Wang, Zhe Wu et al.

To get more accurate saliency maps, recent methods mainly focus on aggregating multi-level features from fully convolutional network (FCN) and introducing edge information as auxiliary supervision. Though remarkable progress has been achieved, we observe that the closer the pixel is to the edge, the more difficult it is to be predicted, because edge pixels have a very imbalance distribution. To address this problem, we propose a label decoupling framework (LDF) which consists of a label decoupling (LD) procedure and a feature interaction network (FIN). LD explicitly decomposes the original saliency map into body map and detail map, where body map concentrates on center areas of objects and detail map focuses on regions around edges. Detail map works better because it involves much more pixels than traditional edge supervision. Different from saliency map, body map discards edge pixels and only pays attention to center areas. This successfully avoids the distraction from edge pixels during training. Therefore, we employ two branches in FIN to deal with body map and detail map respectively. Feature interaction (FI) is designed to fuse the two complementary branches to predict the saliency map, which is then used to refine the two branches again. This iterative refinement is helpful for learning better representations and more precise saliency maps. Comprehensive experiments on six benchmark datasets demonstrate that LDF outperforms state-of-the-art approaches on different evaluation metrics.

Dechao Meng, Liang Li, Xuejing Liu et al.

Vehicle Re-Identification is to find images of the same vehicle from various views in the cross-camera scenario. The main challenges of this task are the large intra-instance distance caused by different views and the subtle inter-instance discrepancy caused by similar vehicles. In this paper, we propose a parsing-based view-aware embedding network (PVEN) to achieve the view-aware feature alignment and enhancement for vehicle ReID. First, we introduce a parsing network to parse a vehicle into four different views, and then align the features by mask average pooling. Such alignment provides a fine-grained representation of the vehicle. Second, in order to enhance the view-aware features, we design a common-visible attention to focus on the common visible views, which not only shortens the distance among intra-instances, but also enlarges the discrepancy of inter-instances. The PVEN helps capture the stable discriminative information of vehicle under different views. The experiments conducted on three datasets show that our model outperforms state-of-the-art methods by a large margin.

Beichen Zhang, Liang Li, Shijie Yang et al.

Active learning is to design label-efficient algorithms by sampling the most representative samples to be labeled by an oracle. In this paper, we propose a state relabeling adversarial active learning model (SRAAL), that leverages both the annotation and the labeled/unlabeled state information for deriving the most informative unlabeled samples. The SRAAL consists of a representation generator and a state discriminator. The generator uses the complementary annotation information with traditional reconstruction information to generate the unified representation of samples, which embeds the semantic into the whole data representation. Then, we design an online uncertainty indicator in the discriminator, which endues unlabeled samples with different importance. As a result, we can select the most informative samples based on the discriminator's predicted state. We also design an algorithm to initialize the labeled pool, which makes subsequent sampling more efficient. The experiments conducted on various datasets show that our model outperforms the previous state-of-art active learning methods and our initially sampling algorithm achieves better performance.

Jun Wei, Shuhui Wang, Qingming Huang

Most of existing salient object detection models have achieved great progress by aggregating multi-level features extracted from convolutional neural networks. However, because of the different receptive fields of different convolutional layers, there exists big differences between features generated by these layers. Common feature fusion strategies (addition or concatenation) ignore these differences and may cause suboptimal solutions. In this paper, we propose the F3Net to solve above problem, which mainly consists of cross feature module (CFM) and cascaded feedback decoder (CFD) trained by minimizing a new pixel position aware loss (PPA). Specifically, CFM aims to selectively aggregate multi-level features. Different from addition and concatenation, CFM adaptively selects complementary components from input features before fusion, which can effectively avoid introducing too much redundant information that may destroy the original features. Besides, CFD adopts a multi-stage feedback mechanism, where features closed to supervision will be introduced to the output of previous layers to supplement them and eliminate the differences between features. These refined features will go through multiple similar iterations before generating the final saliency maps. Furthermore, different from binary cross entropy, the proposed PPA loss doesn't treat pixels equally, which can synthesize the local structure information of a pixel to guide the network to focus more on local details. Hard pixels from boundaries or error-prone parts will be given more attention to emphasize their importance. F3Net is able to segment salient object regions accurately and provide clear local details. Comprehensive experiments on five benchmark datasets demonstrate that F3Net outperforms state-of-the-art approaches on six evaluation metrics.

Xuejing Liu, Liang Li, Shuhui Wang et al.

Weakly supervised referring expression grounding (REG) aims at localizing the referential entity in an image according to linguistic query, where the mapping between the image region (proposal) and the query is unknown in the training stage. In referring expressions, people usually describe a target entity in terms of its relationship with other contextual entities as well as visual attributes. However, previous weakly supervised REG methods rarely pay attention to the relationship between the entities. In this paper, we propose a knowledge-guided pairwise reconstruction network (KPRN), which models the relationship between the target entity (subject) and contextual entity (object) as well as grounds these two entities. Specifically, we first design a knowledge extraction module to guide the proposal selection of subject and object. The prior knowledge is obtained in a specific form of semantic similarities between each proposal and the subject/object. Second, guided by such knowledge, we design the subject and object attention module to construct the subject-object proposal pairs. The subject attention excludes the unrelated proposals from the candidate proposals. The object attention selects the most suitable proposal as the contextual proposal. Third, we introduce a pairwise attention and an adaptive weighting scheme to learn the correspondence between these proposal pairs and the query. Finally, a pairwise reconstruction module is used to measure the grounding for weakly supervised learning. Extensive experiments on four large-scale datasets show our method outperforms existing state-of-the-art methods by a large margin.

Xuejing Liu, Liang Li, Shuhui Wang et al.

Weakly supervised referring expression grounding aims at localizing the referential object in an image according to the linguistic query, where the mapping between the referential object and query is unknown in the training stage. To address this problem, we propose a novel end-to-end adaptive reconstruction network (ARN). It builds the correspondence between image region proposal and query in an adaptive manner: adaptive grounding and collaborative reconstruction. Specifically, we first extract the subject, location and context features to represent the proposals and the query respectively. Then, we design the adaptive grounding module to compute the matching score between each proposal and query by a hierarchical attention model. Finally, based on attention score and proposal features, we reconstruct the input query with a collaborative loss of language reconstruction loss, adaptive reconstruction loss, and attribute classification loss. This adaptive mechanism helps our model to alleviate the variance of different referring expressions. Experiments on four large-scale datasets show ARN outperforms existing state-of-the-art methods by a large margin. Qualitative results demonstrate that the proposed ARN can better handle the situation where multiple objects of a particular category situated together.

Guoli Song, Shuhui Wang, Qingming Huang et al.

Multimodal learning aims to discover the relationship between multiple modalities. It has become an important research topic due to extensive multimodal applications such as cross-modal retrieval. This paper attempts to address the modality heterogeneity problem based on Gaussian process latent variable models (GPLVMs) to represent multimodal data in a common space. Previous multimodal GPLVM extensions generally adopt individual learning schemes on latent representations and kernel hyperparameters, which ignore their intrinsic relationship. To exploit strong complementarity among different modalities and GPLVM components, we develop a novel learning scheme called Harmonization, where latent model parameters are jointly learned from each other. Beyond the correlation fitting or intra-modal structure preservation paradigms widely used in existing studies, the harmonization is derived in a model-driven manner to encourage the agreement between modality-specific GP kernels and the similarity of latent representations. We present a range of multimodal learning models by incorporating the harmonization mechanism into several representative GPLVM-based approaches. Experimental results on four benchmark datasets show that the proposed models outperform the strong baselines for cross-modal retrieval tasks, and that the harmonized multimodal learning method is superior in discovering semantically consistent latent representation.

Junbao Zhuo, Shuhui Wang, Shuhao Cui et al.

We address the unsupervised open domain recognition (UODR) problem, where categories in labeled source domain S is only a subset of those in unlabeled target domain T. The task is to correctly classify all samples in T including known and unknown categories. UODR is challenging due to the domain discrepancy, which becomes even harder to bridge when a large number of unknown categories exist in T. Moreover, the classification rules propagated by graph CNN (GCN) may be distracted by unknown categories and lack generalization capability. To measure the domain discrepancy for asymmetric label space between S and T, we propose Semantic-Guided Matching Discrepancy (SGMD), which first employs instance matching between S and T, and then the discrepancy is measured by a weighted feature distance between matched instances. We further design a limited balance constraint to achieve a more balanced classification output on known and unknown categories. We develop Unsupervised Open Domain Transfer Network (UODTN), which learns both the backbone classification network and GCN jointly by reducing the SGMD, enforcing the limited balance constraint and minimizing the classification loss on S. UODTN better preserves the semantic structure and enforces the consistency between the learned domain invariant visual features and the semantic embeddings. Experimental results show superiority of our method on recognizing images of both known and unknown categories.

Yangyu Chen, Shuhui Wang, Weigang Zhang et al.

In video captioning task, the best practice has been achieved by attention-based models which associate salient visual components with sentences in the video. However, existing study follows a common procedure which includes a frame-level appearance modeling and motion modeling on equal interval frame sampling, which may bring about redundant visual information, sensitivity to content noise and unnecessary computation cost. We propose a plug-and-play PickNet to perform informative frame picking in video captioning. Based on a standard Encoder-Decoder framework, we develop a reinforcement-learning-based procedure to train the network sequentially, where the reward of each frame picking action is designed by maximizing visual diversity and minimizing textual discrepancy. If the candidate is rewarded, it will be selected and the corresponding latent representation of Encoder-Decoder will be updated for future trials. This procedure goes on until the end of the video sequence. Consequently, a compact frame subset can be selected to represent the visual information and perform video captioning without performance degradation. Experiment results shows that our model can use 6-8 frames to achieve competitive performance across popular benchmarks.