Human-object interaction (HOI) detection aims to interpret the interactions of human-object pairs. Existing methods adopt a one-step reasoning paradigm that simultaneously outputs multi-label results for all HOI pairs without distinguishing difficulties. However, there are significant variations among HOI pairs in the same image, making their performance degrade in challenging situations. In this paper, we argue that the model should prioritize hard samples after inferring easy ones, and hard samples can benefit from easy ones. To this end, we propose a novel Multi-step Reasoning Network that progressively learns from easy to hard samples. In particular, an Easy-to-Hard Learning Block is introduced to enhance the representation of hard HOI pairs by prior associations. Additionally, we propose a Multi-step Reasoning Probability Transfer mechanism to enhance multi-label interaction classifications, which leverages cognitive associations and semantic dependencies. Extensive experiments demonstrate that our method outperforms other state-of-the-art on two challenging benchmark datasets.
论文检索
输入标题、作者或关键词，从 7,876 篇学术成果中精准定位
Most previous approaches for analyzing food images have relied on extensively annotated datasets, resulting in significant human labeling expenses due to the varied and intricate nature of such images. Inspired by the effectiveness of contrastive self-supervised methods in utilizing unlabelled data, weiqing explore leveraging these techniques on unlabelled food images. In contrastive self-supervised methods, two views are randomly generated from an image by data augmentations. However, regarding food images, the two views tend to contain similar informative contents, causing large mutual information, which impedes the efficacy of contrastive self-supervised learning. To address this problem, we propose Feature Suppressed Contrast (FeaSC) to reduce mutual information between views. As the similar contents of the two views are salient or highly responsive in the feature map, the proposed FeaSC uses a response-aware scheme to localize salient features in an unsupervised manner. By suppressing some salient features in one view while leaving another contrast view unchanged, the mutual information between the two views is reduced, thereby enhancing the effectiveness of contrast learning for self-supervised food pre-training. As a plug-and-play module, the proposed method consistently improves BYOL and SimSiam by 1.70% ~ 6.69% classification accuracy on four publicly available food recognition datasets. Superior results have also been achieved on downstream segmentation tasks, demonstrating the effectiveness of the proposed method.
Facial expression recognition (FER) plays a critical role in human-computer interaction and affective computing. Traditional FER methods typically rely on comparing the difference between an examined facial expression and a neutral face of the same person to extract the motion of facial features and filter out expression-irrelevant information. With the extensive use of deep learning, the performance of FER has been further improved. However, existing deep learning-based methods rarely utilize neutral faces. To address this gap, we propose a novel deep learning-based FER method called Generative Neutral Features-Disentangled Learning (GNDL), which draws inspiration from the facial feature manifold. Our approach integrates a neutral feature generator (NFG) that generates neutral features in scenarios where the neutral face of the same subject is not available. The NFG uses fine-grained features from examined images as input and produces corresponding neutral features with the same identity. We train the NFG using a neutral feature reconstruction loss to ensure that the generative neutral features are consistent with the actual neutral features. We then disentangle the generative neutral features from the examined features to remove disturbance features and generate an expression deviation embedding for classification. Extensitive experimental results on three popular databases (CK+, Oulu-CASIA, and MMI) demonstrate that our proposed GNDL method outperforms state-of-the-art FER methods.
Vision-language pre-training (VLP) on large-scale datasets has shown premier performance on various downstream tasks. In contrast to plenty of available benchmarks with English corpus, large-scale pre-training datasets and downstream datasets with Chinese corpus remain largely unexplored. In this work, we build a large-scale high-quality Chinese Cross-Modal Benchmark named CCMB for the research community, which contains the currently largest public pre-training dataset Zero and five human-annotated fine-tuning datasets for downstream tasks. Zero contains 250 million images paired with 750 million text descriptions, plus two of the five fine-tuning datasets are also currently the largest ones for Chinese cross-modal downstream tasks. Along with the CCMB, we also develop a VLP framework named R2D2, applying a pre-Ranking + Ranking strategy to learn powerful vision-language representations and a two-way distillation method (i.e., target-guided Distillation and feature-guided Distillation) to further enhance the learning capability. With the Zero and the R2D2 VLP framework, we achieve state-of-the-art performance on twelve downstream datasets from five broad categories of tasks including image-text retrieval, image-text matching, image caption, text-to-image generation, and zero-shot image classification. The datasets, models, and codes are available at https://github.com/yuxie11/R2D2
Despite the broad application of Machine Learning models as a Service (MLaaS), they are vulnerable to model stealing attacks. These attacks can replicate the model functionality by using the black-box query process without any prior knowledge of the target victim model. Existing stealing defenses add deceptive perturbations to the victim's posterior probabilities to mislead the attackers. However, these defenses are now suffering problems of high inference computational overheads and unfavorable trade-offs between benign accuracy and stealing robustness, which challenges the feasibility of deployed models in practice. To address the problems, this paper proposes Isolation and Induction (InI), a novel and effective training framework for model stealing defenses. Instead of deploying auxiliary defense modules that introduce redundant inference time, InI directly trains a defensive model by isolating the adversary's training gradient from the expected gradient, which can effectively reduce the inference computational cost. In contrast to adding perturbations over model predictions that harm the benign accuracy, we train models to produce uninformative outputs against stealing queries, which can induce the adversary to extract little useful knowledge from victim models with minimal impact on the benign performance. Extensive experiments on several visual classification datasets (e.g., MNIST and CIFAR10) demonstrate the superior robustness (up to 48% reduction on stealing accuracy) and speed (up to 25.4× faster) of our InI over other state-of-the-art methods. Our codes can be found in https://github.com/DIG-Beihang/InI-Model-Stealing-Defense.
Cross-network node classification aims to leverage the abundant knowledge from a labeled source network to help classify the node in an unlabeled target network. However, existing methods assume that label sets are identical across domains, which is easily violated in practice. Hence, we attempt to integrate network embedding with universal domain adaptation, which transfers valuable knowledge across domains without assumption on the label sets, to assist in node classification. Nonetheless, the complex network relationships between nodes increase the difficulty of this universal domain adaptive node classification task. In this work, we propose a novel Universal Domain Adaptive Network Embedding (UDANE) framework, which learns transferable node representations across networks to succeed in such a task. Technically, we first adopt the cross-network node embedding component to model comprehensive node information of both networks. Then we employ the inter-domain adaptive alignment component to exploit and relate knowledge across domains, learning domain-invariant representation for knowledge transfer. In addition, the intra-domain contrastive alignment component is proposed to learn discriminative representations beneficial for classification by sufficiently utilizing unlabeled data in the target domain. Extensive experiments have been conducted on real-world datasets, demonstrating that the proposed UDANE model outperforms the state-of-the-art baselines by a large margin.
In recent years, open-vocabulary (OV) object detection has attracted increasing research attention. Unlike traditional detection, which only recognizes fixed-category objects, OV detection aims to detect objects in an open category set. Previous works often leverage vision-language (VL) training data (e.g., referring grounding data) to recognize OV objects. However, they only use pairs of nouns and individual objects in VL data, while these data usually contain much more information, such as scene graphs, which are also crucial for OV detection. In this paper, we propose a novel Scene-Graph-Based Discovery Network (SGDN) that exploits scene graph cues for OV detection. Firstly, a scene-graph-based decoder (SGDecoder) including sparse scene-graph-guided attention (SSGA) is presented. It captures scene graphs and leverages them to discover OV objects. Secondly, we propose scene-graph-based prediction (SGPred), where we build a scene-graph-based offset regression (SGOR) mechanism to enable mutual enhancement between scene graph extraction and object localization. Thirdly, we design a cross-modal learning mechanism in SGPred. It takes scene graphs as bridges to improve the consistency between cross-modal embeddings for OV object classification. Experiments on COCO and LVIS demonstrate the effectiveness of our approach. Moreover, we show the ability of our model for OV scene graph detection, while previous OV scene graph generation methods cannot tackle this task.
Audio-visual zero-shot learning (ZSL) has attracted board attention, as it could classify video data from classes that are not observed during training. However, most of the existing methods are restricted to background scene bias and fewer motion details by employing a single-stream network to process scenes and motion information as a unified entity. In this paper, we address this challenge by proposing a novel dual-stream architecture Motion-Decoupled Spiking Transformer (MDFT) to explicitly decouple the contextual semantic information and highly sparsity dynamic motion information. Specifically, The Recurrent Joint Learning Unit (RJLU) could extract contextual semantic information effectively and understand the environment in which actions occur by capturing joint knowledge between different modalities. By converting RGB images to events, our approach effectively captures motion information while mitigating the influence of background scene biases, leading to more accurate classification results. We utilize the inherent strengths of Spiking Neural Networks (SNNs) to process highly sparsity event data efficiently. Additionally, we introduce a Discrepancy Analysis Block (DAB) to model the audio motion features. To enhance the efficiency of SNNs in extracting dynamic temporal and motion information, we dynamically adjust the threshold of Leaky Integrate-and-Fire (LIF) neurons based on the statistical cues of global motion and contextual semantic information. Our experiments demonstrate the effectiveness of MDFT, which consistently outperforms state-of-the-art methods across mainstream benchmarks. Moreover, we find that motion information serves as a powerful regularization for video networks, where using it improves the accuracy of HM and ZSL by 19.1% and 38.4%, respectively.
Multi-view (representation) learning derives an entity's representation from its multiple observable views to facilitate various downstream tasks. The most challenging topic is how to model unobserved entities and their relationships to specific views. To this end, this work proposes a novel multi-view learning method using a View-Aware parameter Modulation mechanism, termed VAM. The key idea is to use trainable parameters as proxies for unobserved entities and views, such that modeling entity-view relationships is converted into modeling the relationship between proxy parameters. Specifically, we first build a set of trainable parameters to learn a mapping from multi-view data to the unified representation as the entity proxy. Then we learn a prototype for each view and design a Modulation Parameter Generator (MPG) that learns a set of view-aware scale and shift parameters from prototypes to modulate the entity proxy and obtain view proxies. By constraining the representativeness, uniqueness, and simplicity of the proxies and proposing an entity-view contrastive loss, parameters are alternatively updated. We end up with a set of discriminative prototypes, view proxies, and an entity proxy that are flexible enough to yield robust representations for out-of-sample entities. Extensive experiments on five datasets show that the results of our VAM outperform existing methods in both classification and clustering tasks.
Graph classification is a crucial task in many real-world multimedia applications, where graphs can represent various multimedia data types such as images, videos, and social networks. Previous efforts have applied graph neural networks (GNNs) in balanced situations where the class distribution is balanced. However, real-world data typically exhibit long-tailed class distributions, resulting in a bias towards the head classes when using GNNs and limited generalization ability over the tail classes. Recent approaches mainly focus on re-balancing different classes during model training, which fails to explicitly introduce new knowledge and sacrifices the performance of the head classes. To address these drawbacks, we propose a novel framework called Retrieval Augmented Hybrid Network (RAHNet) to jointly learn a robust feature extractor and an unbiased classifier in a decoupled manner. In the feature extractor training stage, we develop a graph retrieval module to search for relevant graphs that directly enrich the intra-class diversity for the tail classes. Moreover, we innovatively optimize a category-centered supervised contrastive loss to obtain discriminative representations, which is more suitable for long-tailed scenarios. In the classifier fine-tuning stage, we balance the classifier weights with two weight regularization techniques, i.e., Max-norm and weight decay. Experiments on various popular benchmarks verify the superiority of the proposed method against state-of-the-art approaches.
As a challenging task, unsupervised person re-identification (Re-ID) aims to optimize the pedestrian matching model based on the unlabeled image frames from surveillance videos. Recently, the fusion with the spatio-temporal clues of pedestrians have been proven effective to improve the performance of classification. However, most of these methods adopt some hard combination approaches by multiplying the visual scores with the spatio-temporal scores, which are sensitive to the noise caused by imprecise estimation of the spatio-temporal patterns in unlabeled datasets and limit the advantage of the fusion model. In this paper, we propose a Graph based Spatio-Temporal Fusion model for high-performance multi-modal person Re-ID, namely G-Fusion, to mitigate the impact of noise. In particular, we construct a graph of pedestrian images by selecting neighboring nodes based on the visual information and the transition time between cameras. Then we use a randomly initialized two-layer GraphSAGE model to obtain the multi-modal affinity matrix between images, and deploy the distillation learning to optimize the visual model by learning the affinity between the nodes. Finally, a graph-based multi-modal re-ranking method is deployed to make the decision in the testing phase for precise person Re-ID. Comprehensive experiments are conducted on two large-scale Re-ID datasets, and the results show that our method achieves a significant improvement of the performance while combined with SOTA unsupervised person Re-ID methods. Specifically, the mAP scores can reach 92.2%, and 80.4% on the Market-1501, and MSMT17 datasets respectively.
Multimodal learning has been widely studied and applied due to its improvement over previous unimodal tasks and its effectiveness on emerging multimodal challenges. However, it has been reported that modal encoders are under-optimized in multimodal learning in contrast to unimodal learning, especially when some modalities are dominant over others. Existing solutions to this problem suffer from two limitations: i) they merely focus on inter-modal balance, failing to consider the influence of intra-modal data on each modality; ii) their implementations heavily rely on unimodal performances or losses, thus being suboptimal for the tasks requiring modal interactions (e.g., visual question answering). To tackle these limitations, we propose I2MCL, a generic Intra- and Inter-Modal Curriculum Learning framework which simultaneously considers both data difficulty and modality balance for multimodal learning. In the intra-modal curriculum, we adopt a pretrained teacher model to obtain knowledge distillation loss as the difficulty measurer, which determines the data weights within the corresponding modality. In the inter-modal curriculum, we utilize a Pareto optimization strategy to measure and compare the gradients from distillation loss and task loss across modalities, capable of determining whether a modality should learn from the task or its teacher. Empirical experiments on various tasks including multimodal classification, visual question answering and visual entailment demonstrate that our proposed I2MCL is able to tackle the under-optimized modality problem and bring consistent improvement to multimodal learning.
Viewport prediction is a crucial aspect of tile-based 360° video streaming system. However, existing trajectory based methods lack of robustness, also oversimplify the process of information construction and fusion between different modality inputs, leading to the error accumulation problem. In this paper, we propose a tile classification based viewport prediction method with Multi-modal Fusion Transformer, namely MFTR. Specifically, MFTR utilizes transformer-based networks to extract the long-range dependencies within each modality, then mine intra- and inter-modality relations to capture the combined impact of user historical inputs and video contents on future viewport selection. In addition, MFTR categorizes future tiles into two categories: user interested or not, and selects future viewport as the region that contains most user interested tiles. Comparing with predicting head trajectories, choosing future viewport based on tile's binary classification results exhibits better robustness and interpretability. To evaluate our proposed MFTR, we conduct extensive experiments on two widely used PVS-HM and Xu-Gaze dataset. MFTR shows superior performance over state-of-the-art methods in terms of average prediction accuracy and overlap ratio, also presents competitive computation efficiency.
With advances in sensing technology, multi-modal data collected from different sources are increasingly available. Multi-modal classification aims to integrate complementary information from multi-modal data to improve model classification performance. However, existing multi-modal classification methods are basically weak in integrating global structural information and providing trustworthy multi-modal fusion, especially in safety-sensitive practical applications (e.g., medical diagnosis). In this paper, we propose a novel Dynamic Poly-attention Network (DPNET) for trustworthy multi-modal classification. Specifically, DPNET has four merits: (i) To capture the intrinsic modality-specific structural information, we design a structure-aware feature aggregation module to learn the corresponding structure-preserved global compact feature representation. (ii) A transparent fusion strategy based on the modality confidence estimation strategy is induced to track information variation within different modalities for dynamical fusion. (iii) To facilitate more effective and efficient multi-modal fusion, we introduce a cross-modal low-rank fusion module to reduce the complexity of tensor-based fusion and activate the implication of different rank-wise features via a rank attention mechanism. (iv) A label confidence estimation module is devised to drive the network to generate more credible confidence. An intra-class attention loss is introduced to supervise the network training. Extensive experiments on four real-world multi-modal biomedical datasets demonstrate that the proposed method achieves competitive performance compared to other state-of-the-art ones.
Audio-Visual Event Localization (AVEL) aims to locate events that are both visible and audible in a video. Existing AVEL methods primarily focus on learning generic localization patterns that are applicable to all events. However, events often exhibit modality biases, such as visual-dominated, audio-dominated, or modality-balanced, which can lead to different localization preferences. These preferences may be overlooked by existing methods, resulting in unsatisfactory localization performance. To address this issue, this paper proposes a novel event-aware localization paradigm, which first identifies the event category and then leverages localization preferences specific to that event for improved event localization. To achieve this, we introduce a memory-assisted metric learning framework, which utilizes historic segments as anchors to adjust the unified representation space for both event classification and event localization. To provide sufficient information for this metric learning, we design a spatial-temporal audio-visual fusion encoder to capture the spatial and temporal interaction between audio and visual modalities. Extensive experiments on the public AVE dataset in both fully-supervised and weakly-supervised settings demonstrate the effectiveness of our approach. Code will be released at https://github.com/ShipingGe/AVEL.
Training multimodal networks requires a vast amount of data due to their larger parameter space compared to unimodal networks. Active learning is a widely used technique for reducing data annotation costs by selecting only those samples that could contribute to improving model performance. However, current active learning strategies are mostly designed for unimodal tasks, and when applied to multimodal data, they often result in biased sample selection from the dominant modality. This unfairness hinders balanced multimodal learning, which is crucial for achieving optimal performance. To address this issue, we propose three guidelines for designing a more balanced multimodal active learning strategy. Following these guidelines, a novel approach is proposed to achieve more fair data selection by modulating the gradient embedding with the dominance degree among modalities. Our studies demonstrate that the proposed method achieves more balanced multimodal learning by avoiding greedy sample selection from the dominant modality. Our approach outperforms existing active learning strategies on a variety of multimodal classification tasks. Overall, our work highlights the importance of balancing sample selection in multimodal active learning and provides a practical solution for achieving more balanced active learning for multimodal classification.
Observing objects' attributes at different levels of detail is a fundamental aspect of how humans perceive and understand the world around them. Existing studies focused on attribute prediction in a flat way, but they overlook the underlying attribute hierarchy, e.g., navy blue is a subcategory of blue. In recent years, large language models, e.g., ChatGPT, have emerged with the ability to perform an extensive range of natural language processing tasks like text generation and classification. The factual knowledge learned by LLM can assist us build the hierarchical relations of visual attributes in the wild. Based on that, we propose a model called the object-specific attribute relation net, which takes advantage of three types of relations among attributes - positive, negative, and hierarchical - to better facilitate attribute recognition in images. Guided by the extracted hierarchical relations, our model can predict attributes from coarse to fine. Additionally, we introduce several evaluation metrics for attribute hierarchy to comprehensively assess the model's ability to comprehend hierarchical relations. Our extensive experiments demonstrate that our proposed hierarchical annotation brings improvements to the model's understanding of hierarchical relations of attributes, and the object-specific attribute relation net can recognize visual attributes more accurately.
Beyond First Impressions: Integrating Joint Multi-modal Cues for Comprehensive 3D Representation
PDF ↗In recent years, 3D representation learning has turned to 2D vision-language pre-trained models to overcome data scarcity challenges. However, existing methods simply transfer 2D alignment strategies, aligning 3D representations with single-view 2D images and coarse-grained parent category text. These approaches introduce information degradation and insufficient synergy issues, leading to performance loss. Information degradation arises from overlooking the fact that a 3D representation should be equivalent to a series of multi-view images and more fine-grained subcategory text. Insufficient synergy neglects the idea that a robust 3D representation should align with the joint vision-language space, rather than independently aligning with each modality. In this paper, we propose a multi-view joint modality modeling approach, termed JM3D, to obtain a unified representation for point cloud, text, and image. Specifically, a novel Structured Multimodal Organizer (SMO) is proposed to address the information degradation issue, which introduces contiguous multi-view images and hierarchical text to enrich the representation of vision and language modalities. A Joint Multi-modal Alignment (JMA) is designed to tackle the insufficient synergy problem, which models the joint modality by incorporating language knowledge into the visual modality. Extensive experiments on ModelNet40 and ScanObjectNN demonstrate the effectiveness of our proposed method, JM3D, which achieves state-of-the-art performance in zero-shot 3D classification. JM3D outperforms ULIP by approximately 4.3% on PointMLP and achieves an improvement of up to 6.5% accuracy on PointNet++ in top-1 accuracy for zero-shot 3D classification on ModelNet40. The source code and trained models for all our experiments are publicly available at https://github.com/Mr-Neko/JM3D.
Audiovisual data is everywhere in this digital age, which raises higher requirements for the deep learning models developed on them. To well handle the information of the multi-modal data is the key to a better audiovisual modal. We observe that these audiovisual data naturally have temporal attributes, such as the time information for each frame in the video. More concretely, such data is inherently multi-modal according to both audio and visual cues, which proceed in a strict chronological order. It indicates that temporal information is important in multi-modal acoustic event modeling for both intra- and inter-modal. However, existing methods deal with each modal feature independently and simply fuse them together, which neglects the mining of temporal relation and thus leads to sub-optimal performance. With this motivation, we propose a Temporal Multi-modal graph learning method for Acoustic event Classification, called TMac, by modeling such temporal information via graph learning techniques. In particular, we construct a temporal graph for each acoustic event, dividing its audio data and video data into multiple segments. Each segment can be considered as a node, and the temporal relationships between nodes can be considered as timestamps on their edges. In this case, we can smoothly capture the dynamic information in intra-modal and inter-modal. Several experiments are conducted to demonstrate TMac outperforms other SOTA models in performance. Our code is available at https://github.com/MGitHubL/TMac.
Incorporating Domain Knowledge Graph into Multimodal Movie Genre Classification with Self-Supervised Attention and Contrastive Learning
PDF ↗Multimodal movie genre classification has always been regarded as a demanding multi-label classification task due to the diversity of multimodal data such as posters, plot summaries, trailers and metadata. Although existing works have made great progress in modeling and combining each modality, they still face three issues: 1) unutilized group relations in metadata, 2) unreliable attention allocation, and 3) indiscriminative fused features. Given that the knowledge graph has been proven to contain rich information, we present a novel framework that exploits the knowledge graph from various perspectives to address the above problems. As a preparation, the metadata is processed into a domain knowledge graph. A translate model for knowledge graph embedding is adopted to capture the relations between entities. Firstly we retrieve the relevant embedding from the knowledge graph by utilizing group relations in metadata and then integrate it with other modalities. Next, we introduce an Attention Teacher module for reliable attention allocation based on self-supervised learning. It learns the distribution of the knowledge graph and produces rational attention weights. Finally, a Genre-Centroid Anchored Contrastive Learning module is proposed to strengthen the discriminative ability of fused features. The embedding space of anchors is initialized from the genre entities in the knowledge graph. To verify the effectiveness of our framework, we collect a larger and more challenging dataset named MM-IMDb 2.0 compared with the MM-IMDb dataset. The experimental results on two datasets demonstrate that our model is superior to the state-of-the-art methods. Our code and dataset is available at https://github.com/aoluming/IDKG.git.