ACM MM 2024 Papers

Large vision-language models (LVLMs) integrate visual information into large language models, showcasing remarkable multi-modal conversational capabilities. However, the visual modules introduces new challenges in terms of robustness for LVLMs, as attackers can craft adversarial images that are visually clean but may mislead the model to generate incorrect answers. In general, LVLMs rely on vision encoders to transform images into visual tokens, which are crucial for the language models to perceive image contents effectively. Therefore, we are curious about one question: Can LVLMs still generate correct responses when the encoded visual tokens are attacked and disrupting the visual information? To this end, we propose a non-targeted attack method referred to as VT-Attack (Visual Tokens Attack), which constructs adversarial examples from multiple perspectives, with the goal of comprehensively disrupting feature representations and inherent relationships as well as the semantic properties of visual tokens output by image encoders. Using only access to the image encoder in the proposed attack, the generated adversarial examples exhibit transferability across diverse LVLMs utilizing the same image encoder and generality across different tasks. Extensive experiments validate the superior attack performance of the VT-Attack over baseline methods, demonstrating its effectiveness in attacking LVLMs with image encoders, which in turn can provide guidance on the robustness of LVLMs, particularly in terms of the stability of the visual feature space.

Few-Shot Learning (FSL) aims to recognize new classes with limited labeled data. Recent studies have attempted to address the challenge of rare samples with textual prompts to modulate visual features. However, they usually struggle to capture complex semantic relationships between textual and visual features. Moreover, vanilla self-attention is heavily affected by useless information in images, severely constraining the potential of semantic priors in FSL due to the confusion of numerous irrelevant tokens during interaction. To address these aforementioned issues, a K-NN Transformer with Pyramid Prompts (KTPP) is proposed to select discriminative information with K-NN Context Attention (KCA) and adaptively modulate visual features with Pyramid Cross-modal Prompts (PCP). First, for each token, the KCA only selects the K most relevant tokens to compute the self-attention matrix and incorporates the mean of all tokens as the context prompt to provide the global context in three cascaded stages. As a result, irrelevant tokens can be progressively suppressed. Secondly, pyramid prompts are introduced in the PCP to emphasize visual features via interactions between text-based class-aware prompts and multi-scale visual features. This allows the ViT to dynamically adjust the importance weights of visual features based on rich semantic information at different scales, making models robust to spatial variations. Finally, augmented visual features and class-aware prompts are interacted via the KCA to extract class-specific features. Consequently, our model further enhances noise-free visual representations via deep cross-modal interactions, extracting generalized visual representation in scenarios with few labeled samples. Extensive experiments on four benchmark datasets demonstrate significant gains over the state-of-the-art methods, especially for the 1-shot task with 2.28% improvement on average due to semantically enhanced visual representations.

Single hyperspectral image super-resolution aims to reconstruct a high-resolution hyperspectral image (HRHSI) from an observed low resolution hyperspectral image (LRHSI). Most current methods combine CNN and Transformer structures to directly extract features of all channels in LRHSI for image reconstruction, but they do not consider the interference of redundant information in adjacent bands, resulting in spectral and spatial distortions in the reconstruction results and an increase in model computational complexity. To address this issue, this paper proposes a spectral clustering-based pyramid super-resolution network (SCPSN) to progressively reconstruct HRHSI at different scales. In each image reconstruction layer, a clustering super-resolution block (CSRB) consisting of spectral clustering block (SCB), patch non local attention block (PNAB), and dynamic fusion block (DFB) is designed to achieve the reconstruction of detail features. Specifically, for the high correlation between adjacent spectral bands in LRHSI, a SCB is first constructed to achieve clustering of spectral channels and filtering of hyperchannels. This can reduce the interference of redundant spectral information and the computational complexity of the model. Then, by utilizing the non-local similarity of features within the channel, a patch non-local attention block (PNAB) is constructed to enhance the features of hyperchannels. Next, a dynamic fusion block (DFB) is designed to reconstruct the features of all channels in LRHSI by establishing correlations between enhanced hyperchannels and other channels. Finally, the reconstructed channels are upsampled and added to the corresponding channels to obtain the reconstructed HRHSI. Extensive experiments validate that the performance of SCPSN is superior to that of some other state-of-the-art (SOTA) HSSR methods in terms of visual effects and quantitative metrics. In addition, our model does not require training on large-scale datasets compared to other methods. The dataset and code will be released on GitHub.

Building a unified model for general low-level vision tasks holds significant research and practical value. Current methods encounter several critical issues. Multi-task restoration approaches can address multiple degradation-to-clean restoration tasks, while their applicability to tasks with different target domains (e.g., image stylization) is limited. Methods like PromptGIP can handle multiple input-target domains but rely on the Masked Autoencoder (MAE) paradigm. Consequently, they are tied to the ViT architecture, resulting in suboptimal image reconstruction quality. In addition, these methods are sensitive to prompt image content and often struggle with low-frequency information processing. In this paper, we propose a Visual task Prompt-based Image Processing (VPIP) framework to overcome these challenges. VPIP employs visual task prompts to manage tasks with different input-target domains and allows flexible selection of backbone network suitable for general tasks. Besides, a new prompt cross-attention is introduced to facilitate interaction between the input and prompt information. Based on the VPIP framework, we train a low-level vision generalist model, namely GenLV, on 30 diverse tasks. Experimental results show that GenLV can successfully address a variety of low-level tasks, significantly outperforming existing methods both quantitatively and qualitatively. Codes are available at https://github.com/chxy95/GenLV.

Since the release of the CLIP model by OpenAI, it has received widespread attention. However, categories in the real world often exhibit a long-tail distribution, and existing CLIP models struggle to effectively recognize rare, tail-end classes, such as an endangered African bird. An intuitive idea is to generate visual descriptions for these tail-end classes and use descriptions to create category prototypes for classification. However, experiments reveal that visual descriptions, image captions, and test prompt templates belong to three distinct domains, leading to distribution shifts. In this paper, we propose the use of caption object parsing to identify the objects set contained within captions. During training, the object sets is used to generate visual descriptions and test prompts, aligning these three domains and enabling the text encoder to generate category prototypes based on visual descriptions. Thanks to the acquired object sets, our approach can construct many-to-many relationships at a lower cost and derive soft labels, addressing the noise issues associated with traditional one-to-one matching. Extensive experimental results demonstrate that our method significantly surpasses the CLIP baseline and exceeds existing methods, achieving a new state-of-the-art (SOTA).

Point clouds from real-world scenarios inevitably contain complex noise, significantly impairing the accuracy of downstream tasks. To tackle this challenge, cascading encoder-decoder architecture has become a conventional technical route to iterative denoise. However, circularly feeding the output of denoiser as its input again involves the re-extraction of underlying surface, leading to unstable denoising process and over-smoothed geometric details. To address these issues, we propose a novel denoising paradigm dubbed PD-Refiner that employs a single encoder to model the underlying surface. Then, we leverage several lightweight hierarchical Underlying Surface Inheritance Refiners (USIRs) to inherit and strengthen it, thereby avoiding the re-extraction from the intermediate point cloud. Furthermore, we design adaptive edge-aware supervision to improve the edge awareness of the USIRs, allowing for the adjustment of the denoising preferences from global structure to local details. The results demonstrate that our method not only achieves state-of-the-art performance in terms of denoising stability and efficacy, but also enhances edge clarity and point cloud uniformity.

Many domain adaptive object detection (DAOD) methods employ domain adversarial training to align features and mitigate the domain gap. In this approach, a feature extractor is trained to deceive a domain classifier, thereby aligning feature distributions. However, the domain classifier's discrimination capability can easily fall into a local optimum due to the equilibrium challenge, hindering the effective training of the feature extractor. In this work, we propose an efficient optimization strategy called Virtual-label Fooled Domain Discrimination (VFDD), which revitalizes the domain classifier during training using virtual domain labels. Such virtual label makes the separable distributions less separable, and thus leads to a more easily confused domain classifier, which in turn further drives feature alignment. Particularly, we introduce a novel concept of virtual domain label for the unaligned samples and propose the VirtualH -divergence to overcome the problem of falling into local optimum due to the equilibrium challenge. VFDD is orthogonal to most existing DAOD methods and can be integrated as a plug-and-play module to enhance these models. Theoretical insights and experimental analyses demonstrate that VFDD improves many popular baselines and surpasses recent unsupervised DAOD models.

Referring image segmentation (RIS) aims to segment a particular region based on a specific expression. Existing one-stage methods have explored various fusion strategies, yet they encounter two significant issues. Primarily, most methods rely on manually selected visual features from the visual encoder layers. Moreover, the direct fusion of word-level features into coarse aligned features disrupts the established vision-language alignment. In this paper, we introduce an innovative framework for RIS that seeks to overcome these challenges with adaptive alignment of vision and language features, termed the Adaptive Selection with Dual Alignment (ASDA). ASDA innovates in two aspects. Firstly, we design an Adaptive Feature Selection and Fusion (AFSF) module to dynamically select visual features focusing on different regions related to various descriptions. AFSF is equipped with scale-wise feature aggregator to provide hierarchically coarse features that preserve crucial low-level details. Secondly, a Word Guided Dual-Branch Aligner (WGDA) is leveraged to integrate coarse features with linguistic cues by word-guided attention, which effectively addresses the common issue of vision-language misalignment. Extensive experimental results demonstrate that our ASDA framework surpasses state-of-the-art methods on RefCOCO, RefCOCO+ and G-Ref benchmark.

Domain generalization (DG) task aims to learn a robust model from source domains that could handle the out-of-distribution (OOD) issue. In order to improve the generalization ability of the model in unseen domains, increasing the diversity of training samples is an effective solution. However, existing augmentation approaches always have some limitations. On the one hand, the augmentation manner in most DG methods is not enough as the model may not see the perturbed features in approximate the worst case due to the randomness, thus the transferability in features could not be fully explored. On the other hand, the causality in discriminative features is not involved in these methods, which harms the generalization ability of model due to the spurious correlations. To address these issues, we propose a Dual-stream Feature Augmentation~(DFA) method by constructing some hard features from two perspectives. Firstly, to improve the transferability, we construct some targeted features with domain related augmentation manner. Through the guidance of uncertainty, some hard cross-domain fictitious features are generated to simulate domain shift. Secondly, to take the causality into consideration, the spurious correlated non-causal information is disentangled by an adversarial mask, then the more discriminative features can be extracted through these hard causal related information. Different from previous fixed synthesizing strategy, the two augmentations are integrated into a unified learnable feature disentangle model. Based on these hard features, contrastive learning is employed to keep the semantic consistency and improve the robustness of the model. Extensive experiments on several datasets demonstrated that our approach could achieve state-of-the-art performance for domain generalization. Our code is available at: https://github.com/alusi123/DFA.

Neural radiance fields are capable of reconstructing high-quality drivable human avatars but are expensive to train and render and not suitable for multi-human scenes with complex shadows. To reduce consumption, we propose Animatable 3D Gaussian, which learns human avatars from input images and poses. We extend 3D Gaussians to dynamic human scenes by modeling a set of skinned 3D Gaussians and a corresponding skeleton in canonical space and deforming 3D Gaussians to posed space according to the input poses. We introduce a multi-head hash encoder for pose-dependent shape and appearance and a time-dependent ambient occlusion module to achieve high-quality reconstructions in scenes containing complex motions and dynamic shadows. On both novel view synthesis and novel pose synthesis tasks, our method achieves higher reconstruction quality than InstantAvatar with less training time (1/60), less GPU memory (1/4), and faster rendering speed (7x). Our method can be easily extended to multi-human scenes and achieve comparable novel view synthesis results on a scene with ten people in only 25 seconds of training.

Cloth-Changing Person Re-Identification (CC-ReID) aims to accurately identify a target person in the more realistic surveillance scenario where clothes of the pedestrian may change drastically, which is critical in public security systems for tracking down disguised criminal suspects. Existing methods mainly transform the CC-ReID problem into cross-modality feature alignment from the data-driven perspective, without modelling the interference factors such as clothes and camera view changes meticulously. This may lead to over-consideration or under-consideration of the influence of these factors on the extraction of robust and discriminative identity features. This paper proposes a novel algorithm for thoroughly disentangling identity features from interference factors brought by clothes and camera view changes while ensuring the robustness and discriminability. It adopts a dual-stream identity feature learning framework consisting of a raw image stream and a cloth-erasing stream, to explore discriminative and cloth-irrelevant identity feature representations. Specifically, an adaptive cloth-irrelevant contrastive objective is introduced to contrast features extracted by the two streams, aiming to suppress the fluctuation caused by clothes textures in the identity feature space. Moreover, we innovatively mitigate the influence of the interference factors through a generative adversarial interference factor decoupling network. This network is targeted at capturing identity-related information residing in the interference factors and disentangling the identity features from such information. Extensive experimental results demonstrate the effectiveness of the proposed method, achieving superior performances to state-of-the-art methods.

Multi-view clustering (MVC) methods based on non-negative matrix factorization (NMF) have gained popularity owing to their ability to provide interpretable clustering results. However, these NMF-based MVC methods generally process each view independently and thus ignore the potential relationship between views. Besides, they are limited in the ability to capture nonlinear data structures. To overcome these weaknesses and inspired by deep learning, we propose a multi-view clustering method based on deep non-negative tensor factorization (MVC-DNTF). With deep tensor factorization, our method can well exploit the spatial structure of the original data and is capable of extracting more deep and nonlinear features embedded in different views. To further extract the complementary information of different views, we adopt the weighted tensor Schatten p-norm regularization term. An optimization algorithm is developed to effectively solve the MVC-DNTF objective. Extensive experiments are performed to demonstrate the effectiveness and superiority of our method.

We are investigating interaction with entities showing features different from humans', to understand how they can be embodied as avatars and perceived as living, social beings. To push this investigation to its limit, we have designed as an avatar an interactive space (the Room), that challenges both the anthropomorphic structure, and most of the social interaction mechanisms we are used to. We introduce a first framework for the Room design, addressing challenges related to its body, perception, and interaction process. We present a pilot implementation of some of the aspects of the framework as an interactive installation, namely a real-time, two-player, VR experience, featuring the Room avatar, with a focus on haptic feedback as the main means of perception for the subject embodying the Room. By radically challenging anthropomorphism, we seek to investigate the most basic aspects of embodiment and social cognition.

In this work, we present AerialGait, a comprehensive dataset for aerial-ground gait recognition. This dataset comprises 82,454 sequences totaling over 10 million frames from 533 subjects, captured from both aerial and ground perspectives. To align with real-life scenarios of aerial and ground surveillance, we utilize a drone and a ground surveillance camera for data acquisition. The drone is operated at various speeds, directions, and altitudes. Meanwhile, we conduct data collection across five diverse surveillance sites to ensure a comprehensive simulation of real-world settings. AerialGait has several unique features: 1) The gait sequences exhibit significant variations in views, resolutions, and illumination across five distinct scenes. 2) It incorporates challenges of motion blur and frame discontinuity due to drone mobility. 3) The dataset reflects the domain gap caused by the view disparity between aerial and ground views, presenting a realistic challenge for drone-based gait recognition. Moreover, we perform a comprehensive analysis of existing gait recognition methods on AerialGait dataset and propose the Aerial-Ground Gait Network (AGG-Net). AGG-Net effectively learns discriminative features from aerial views by uncertainty learning and clusters features across aerial and ground views through prototype learning. Our model achieves state-of-the-art performance on both AerialGait and DroneGait datasets.

Nowadays, misinformation is widely spreading over various social media platforms and causes extremely negative impacts on society. To combat this issue, automatically identifying misinformation, especially those containing multimodal content, has attracted growing attention from the academic and industrial communities, and induced an active research topic named Multimodal Misinformation Detection (MMD). Typically, existing MMD methods capture the semantic correlation and inconsistency between multiple modalities, but neglect some potential clues in multimodal content. Recent studies suggest that manipulated traces of the images in articles are non-trivial clues for detecting misinformation. Meanwhile, we find that the underlying intentions behind the manipulation, e.g., harmful and harmless, also matter in MMD. Accordingly, in this work, we propose to detect misinformation by learning manipulation features that indicate whether the image has been manipulated, as well as intention features regarding the harmful and harmless intentions of the manipulation. Unfortunately, the manipulation and intention labels that make these features discriminative are unknown. To overcome the problem, we propose two weakly supervised signals as alternatives by introducing additional datasets on image manipulation detection and formulating two classification tasks as positive and unlabeled learning problems. Based on these ideas, we propose a novel MMD method, namely Harmfully Manipulated Images Matter in MMD (Hami-m3d). Extensive experiments across three benchmark datasets can demonstrate that Hami-m3d can consistently improve the performance of any MMD baselines.

Multimodal Relation Extraction (MRE) has achieved great improvements. However, modern MRE models are easily affected by irrelevant objects during multimodal alignment which are called error sensitivity issues. The main reason is that visual features are not fully aligned with textual features and the reasoning process may suppress redundant and noisy information at the risk of losing critical information. In light of this, we propose a Caption-Aware Multimodal Relation Extraction Network with Mutual Information Maximization (CAMIM). Specifically, we first generate detailed image captions through the Large Language Model (LLM). Then, the Caption-Aware Module (CAM) hierarchically aligns the fine-grained visual entities and textual entities for reasoning. In addition, for preserving crucial information within different modalities, we leverage a Mutual Information Maximization method to regulate the multimodal reasoning module. Experiments show that our model outperforms the state-of-the-art MRE models on the benchmark dataset MNRE. Further ablation studies prove the pluggable and effective performance of our Caption-Aware Module and Mutual Information Maximization method. Our code is available at https://github.com/zefanZhang-cn/CAMIM.

Quality assessment and aesthetics assessment aim to evaluate the perceived quality and aesthetics of visual content. Current learning-based methods suffer greatly from the scarcity of labeled data and usually perform sub-optimally in terms of generalization. Although masked image modeling (MIM) has achieved noteworthy advancements across various high-level tasks (e.g., classification, detection etc.). In this work, we take on a novel perspective to investigate its capabilities in terms of quality- and aesthetics-awareness. To this end, we propose Quality- and aesthetics-aware pretraining (QPT V2), the first pretraining framework based on MIM that offers a unified solution to quality and aesthetics assessment. To perceive the high-level semantics and fine-grained details, pretraining data is curated. To comprehensively encompass quality- and aesthetics-related factors, degradation is introduced. To capture multi-scale quality and aesthetic information, model structure is modified. Extensive experimental results on 11 downstream benchmarks clearly show the superior performance of QPT V2 in comparison with current state-of-the-art approaches and other pretraining paradigms. Code and models will be released at https://github.com/KeiChiTse/QPT-V2.