ACM MM 2024 论文列表

Skeleton-based action recognition has garnered significant attention due to the utilization of concise and resilient skeletons. Nevertheless, the absence of detailed body information in skeletons restricts performance, while other multimodal methods require substantial inference resources and are inefficient when using multimodal data during both training and inference stages. To address this and fully harness the complementary multimodal features, we propose a novel multi-modality co-learning (MMCL) framework by leveraging the multimodal large language models (LLMs) as auxiliary networks for efficient skeleton-based action recognition, which engages in multi-modality co-learning during the training stage and keeps efficiency by employing only concise skeletons in inference. Our MMCL framework primarily consists of two modules. First, the Feature Alignment Module (FAM) extracts rich RGB features from video frames and aligns them with global skeleton features via contrastive learning. Second, the Feature Refinement Module (FRM) uses RGB images with temporal information and text instruction to generate instructive features based on the powerful generalization of multimodal LLMs. These instructive text features will further refine the classification scores and the refined scores will enhance the model's robustness and generalization in a manner similar to soft labels. Extensive experiments on NTU RGB+D, NTU RGB+D 120 and Northwestern-UCLA benchmarks consistently verify the effectiveness of our MMCL, which outperforms the existing skeleton-based action recognition methods. Meanwhile, experiments on UTD-MHAD and SYSU-Action datasets demonstrate the commendable generalization of our MMCL in zero-shot and domain-adaptive action recognition. Our code is publicly available at: https://github.com/liujf69/MMCL-Action.

Recently, text-to-3D generation has attracted significant attention, resulting in notable performance enhancements. Previous methods utilize end-to-end 3D generation models to initialize 3D Gaussians, multi-view diffusion models to enforce multi-view consistency, and text-to-image diffusion models to refine details with score distillation algorithms. However, these methods exhibit two limitations. Firstly, they encounter conflicts in generation directions since different models aim to produce diverse 3D assets. Secondly, the issue of over-saturation in score distillation has not been thoroughly investigated and solved. To address these limitations, we propose PlacidDreamer, a text-to-3D framework that harmonizes initialization, multi-view generation, and text-conditioned generation with a single multi-view diffusion model, while simultaneously employing a novel score distillation algorithm to achieve balanced saturation. To unify the generation direction, we introduce the Latent-Plane module, a training-friendly plug-in extension that enables multi-view diffusion models to provide fast geometry reconstruction for initialization and enhanced multi-view images to personalize the text-to-image diffusion model. To address the over-saturation problem, we propose to view score distillation as a multi-objective optimization problem and introduce the Balanced Score Distillation algorithm, which offers a Pareto Optimal solution that achieves both rich details and balanced saturation. Extensive experiments validate the outstanding capabilities of our PlacidDreamer. The code is available at https://github.com/HansenHuang0823/PlacidDreamer.

Feature upsampling is an essential operation in constructing deep convolutional neural networks. However, existing upsamplers either lack specific feature guidance or necessitate the utilization of high-resolution feature maps, resulting in a loss of performance and flexibility. In this paper, we find that the local self-attention naturally has the feature guidance capability, and its computational paradigm aligns closely with the essence of feature upsampling (i.e. feature reassembly of neighboring points). Therefore, we introduce local self-attention into the upsampling task and demonstrate that the majority of existing upsamplers can be regarded as special cases of upsamplers based on local self-attention. Considering the potential semantic gap between upsampled points and their neighboring points, we further introduce the deformation mechanism into the upsampler based on local self-attention, thereby proposing LDA-AQU. As a novel dynamic kernel-based upsampler, LDA-AQU utilizes the feature of queries to guide the model in adaptively adjusting the position and aggregation weight of neighboring points, thereby meeting the upsampling requirements across various complex scenarios. In addition, LDA-AQU is lightweight and can be easily integrated into various model architectures. We evaluate the effectiveness of LDA-AQU across four dense prediction tasks: object detection, instance segmentation, panoptic segmentation, and semantic segmentation. LDA-AQU consistently outperforms previous state-of-the-art upsamplers, achieving performance enhancements of 1.7 AP, 1.5 AP, 2.0 PQ, and 2.5 mIoU compared to the baseline models in the aforementioned four tasks, respectively.

Portrait video editing has attracted wide attention thanks to its practical applications. Existing methods either target fixed-length clips or perform temporally inconsistent per-frame editing. In this work, we present a brand new system, StreamEdit, which is primarily designed to edit streaming videos. Our system follows the ideology of editing propagation to ensure temporal consistency. Concretely, we choose to edit only one reference frame and warp the outcome to obtain the editing results of other frames. For this purpose, we employ a warping module, aided by a probabilistic pixel correspondence estimation network, to help establish the pixel-wise mapping between two frames. However, such a pipeline requires the reference frame to contain all contents appearing in the video, which is scarcely possible especially when there exist large motions and occlusions. To address this challenge, we propose to adaptively replace the reference frame, benefiting from a heuristic strategy referring to the overall pixel mapping uncertainty. That way, we can easily align the editing of the before- and after-replacement reference frames via image inpainting. Extensive experimental results demonstrate the effectiveness and generalizability of our approach in editing streaming portrait videos. Code will be made public.

Recently, large vision language models (LVLMs) have advanced AI by integrating visual and linguistic data for tasks like visual conversation, image captioning, and visual question answering. Current LVLM research either scales up model size for performance or reduces parameters for limited computational resources. We believe both large and tiny models have unique strengths and that collaborative training yields better results than independent training. We propose Collaborative Training of Tiny-Large Vision Language Models (CTVLMs), a framework connecting large and tiny models via a projection layer and leveraging a synergistic training strategy. Our framework improves training efficiency by strengthening the interconnection between large and tiny models. Using the parameter efficiency of tiny models, we effectively align image-text features, then apply knowledge distillation to help large models better align cross-modal information. During fine-tuning, the large model's extensive knowledge enhances tiny model's performance. This collaborative approach allows models to adapt to various computational resources and outperforms existing methods in vision-language tasks.

With the increased attention to model efficiency, post-training sparsity (PTS) has become more and more prevalent because of its effectiveness and efficiency. However, there remain questions on better practice of PTS algorithms and the sparsification ability of models, which hinders the further development of this area.Therefore, a benchmark to comprehensively investigate the issues above is urgently needed. In this paper, we propose the first comprehensive post-training sparsity benchmark called PTSBench towards algorithms and models. We benchmark 10+ PTS general-pluggable fine-grained techniques on 3 typical tasks using over 40 off-the-shelf model architectures. Through extensive experiments and analyses, we obtain valuable conclusions and provide several insights from both algorithms and model aspects. Our PTSBench can provide (1) new observations for a better understanding of the PTS algorithms, (2) in-depth and comprehensive evaluations for the sparsification ability of models, and (3) a well-structured and easy-integrate open-source framework. We hope this work will provide illuminating conclusions and advice for future studies of post-training sparsity methods and sparsification-friendly model design. The code for our PTSBench is released at https://github.com/ModelTC/msbench.

Voice is one of the most widely used media for information transmission in human society. While high-quality synthetic voices are extensively utilized in various applications, they pose significant risks to content security and trust building. Numerous studies have concentrated on AI-synthesized voice detection to mitigate these risks, with many claiming to achieve promising performance. However, recent research has demonstrated that fake voice detectors suffer from serious overfitting to speaker-irrelative features (SiFs) and cannot be used in real-world scenarios. In this paper, we analyze the limitations of existing fake voice detectors and propose a new design philosophy, guiding the detection model to prioritize learning human voice features rather than the difference between the human voice and the synthetic voice. Based on this philosophy, we propose a novel AI-synthesized voice detection framework named SiFSafer, which uses pre-trained speech representation models to enhance the learning of feature distribution in human voices and the adapter fine-tuning to optimize the performance. The evaluation shows that the average EERs of existing fake voice detectors in the ASVspoof datasets can exceed 20% if the SiFs like silence segments are removed, while SiFSafer achieves an EER of less than 8%, indicating that SiFSafer is robust to SiFs and strongly resistant to existing attacks.

In this work, we introduce a novel approach to single-source domain generalization (SDG) in medical imaging, focusing on overcoming the challenge of style variation in out-of-distribution (OOD) domains without requiring domain labels or additional generative models. We propose a Universal Frequency Perturbation framework for SDG termed as UniFreqSDG, that performs hierarchical feature-level frequency domain perturbations, facilitating the model's ability to handle diverse OOD styles. Specifically, we design a learnable spectral perturbation module that adaptively learns the frequency distribution range of samples, allowing for precise low-frequency (LF) perturbation. This adaptive approach not only generates stylistically diverse samples but also preserves domain-invariant anatomical features without the need for manual hyperparameter tuning. Then, the frequency features before and after perturbation are decoupled and recombined through the Content Preservation Reconstruction operation, effectively preventing the loss of discriminative content information. Furthermore, we introduce the Active Domain-variance Inducement Loss to encourage effective perturbation in the frequency domain while ensuring the sufficient decoupling of domain-invariant and domain-style features. Extensive experiments demonstrate that UniFreqSDG increases the dice score by an average of 7.47% (from 77.98% to 85.45%) on the fundus dataset and 4.99% (from 71.42% to 76.73%) on the prostate dataset compared to the state-of-the-art approaches.

Fully test-time adaptation aims to adapt a network model online based on sequential analysis of input samples during the inference stage. We observe that, when applying a transformer network model into a new domain, the self-attention profiles of image samples in the target domain deviate significantly from those in the source domain, which results in large performance degradation during domain changes. To address this important issue, we propose a new structure for the self-attention modules in the transformer. Specifically, we incorporate three domain-conditioning vectors, called domain conditioners, into the query, key, and value components of the self-attention module. We learn a network to generate these three domain conditioners from the class token at each transformer network layer. We find that, during fully online test-time adaptation, these domain conditioners at each transform network layer are able to gradually remove the impact of domain shift and largely recover the original self-attention profile. Our extensive experimental results demonstrate that the proposed domain-conditioned transformer significantly improves the online fully test-time domain adaptation performance and outperforms existing state-of-the-art methods by large margins.

Medical image segmentation is of great significance to disease diagnosis and treatment planning. Despite multiple progresses, most present methods (1) pay insufficient attention to suppressing background noise disturbance that impacts segmentation accuracy and (2) are not efficient enough, especially when the images are of large resolutions. To address the two challenges, we turn to a traditional de-noising method and a new efficient network structure and propose BSBP-RWKV for accurate and efficient medical image segmentation. Specifically, we combine the advantages of Perona-Malik Diffusion (PMD) in noise suppression without losing boundary details and RWKV in its efficient structure, and devise the DWT-PMD RWKV Block across one of our encoder branches to preserve boundary details of lesion areas while suppressing background noise disturbance in an efficient structure. Then we feed the de-noised lesion boundary cues to our proposed Multi-Step Runge-Kutta convolutional Block to supplement the cues with more local details. We also propose a novel loss function for shape refinement that can align the shape of predicted lesion areas with GT masks in both spatial and frequency domains. Experiments on ISIC 2016 and Kvasir-SEG show the superior accuracy and efficiency of our BSBP-RWKV. Specifically, BSBP-RWKV reduces complexity of 5.8 times compared with the SOTA while also cutting down GPU memory usage by over 62.7% for each 1024 \times 1024 image during inference.

Artwork analysis is important and fundamental skill for art appreciation, which could enrich personal aesthetic sensibility and facilitate the critical thinking ability. Understanding artworks is challenging due to its subjective nature, diverse interpretations, and complex visual elements, requiring expertise in art history, cultural background, and aesthetic theory. However, limited by the data collection and model ability, previous works for automatically analyzing artworks mainly focus on classification, retrieval, and other simple tasks, which is far from the goal of AI. To facilitate the research progress, in this paper, we step further to compose comprehensive analysis inspired by the remarkable perception and generation ability of large multimodal models. Specifically, we first propose a task of composing paragraph analysis for artworks, i.e., painting in this paper, only focusing on visual characteristics to formulate more comprehensive understanding of artworks. To support the research on formal analysis, we collect a large dataset PaintingForm, with about 19k painting images and 50k analysis paragraphs. We further introduce a superior large multimodal model for painting analysis composing, dubbed GalleryGPT, which is slightly modified and fine-tuned based on LLaVA architecture leveraging our collected data. We conduct formal analysis generation and zero-shot experiments across several datasets to assess the capacity of our model. The results show remarkable performance improvements comparing with powerful baseline LMMs, demonstrating its superb ability of art analysis and generalization. \textcolor{blue}{The codes and model are available at: https://github.com/steven640pixel/GalleryGPT.

Recently, the 3D Gaussian Splatting (3D-GS) method has achieved great success in novel view synthesis, providing real-time rendering while ensuring high-quality rendering results. However, this method faces challenges in modeling specular reflections and handling anisotropic appearance components, especially in dealing with view-dependent color under complex lighting conditions. Additionally, 3D-GS uses spherical harmonic to learn the color representation, which has limited ability to represent complex scenes. To overcome these challenges, we introduce Lantent-SpecGS, an approach that utilizes a universal latent neural descriptor within each 3D Gaussian. This enables a more effective representation of 3D feature fields, including appearance and geometry. Moreover, two parallel CNNs are designed to decoder the splatting feature maps into diffuse color and specular color separately. A mask that depends on the viewpoint is learned to merge these two colors, resulting in the final rendered image. Experimental results demonstrate that our method obtains competitive performance in novel view synthesis and extends the ability of 3D-GS to handle intricate scenarios with specular reflections.

Cross-modal hashing has emerged as a promising technique for retrieving relevant information across distinct media types thanks to its low storage cost and high retrieval efficiency. However, the success of most existing methods heavily relies on large-scale well-annotated datasets, which are costly and scarce in the real world due to ubiquitous labeling noise. To tackle this problem, in this paper, we propose a novel framework, termed Noise Resistance Cross-modal Hashing (NRCH), to learn hashing with noisy labels by overcoming two key challenges, i.e. noise overfitting and error accumulation. Specifically, i) to mitigate the overfitting issue caused by noisy labels, we present a novel Robust Contrastive Hashing loss (RCH) to target homologous pairs instead of noisy positive pairs, thus avoiding overemphasizing noise. In other words, RCH enforces the model focus on more reliable positives instead of unreliable ones constructed by noisy labels, thereby enhancing the robustness of the model against noise; ii) to circumvent error accumulation, a Dynamic Noise Separator (DNS) is proposed to dynamically and accurately separate the clean and noisy samples by adaptively fitting the loss distribution, thus alleviate the adverse influence of noise on iterative training. Finally, we conduct extensive experiments on four widely used benchmarks to demonstrate the robustness of our NRCH against noisy labels for cross-modal retrieval. The code is available at: https://github.com/LonganWANG-cs/NRCH.git.

Graph neural networks (GNNs) have a wide range of applications in multimedia. Recent studies have shown that Graph neural networks (GNNs) are vulnerable to link stealing attacks, which infers the existence of edges in the target GNN's training graph. Existing attacks are usually based on the assumption that links exist between two nodes that share similar posteriors; however, they fail to focus on links that do not hold under this assumption. To this end, we propose LinkThief, an improved link stealing attack that combines generalized structure knowledge with node similarity, in a scenario where the attackers' background knowledge contains partially leaked target graph and shadow graph. Specifically, to equip the attack model with insights into the link structure spanning both the shadow graph and the target graph, we introduce the idea of creating a Shadow-Target Bridge Graph and extracting edge subgraph structure features from it. Through theoretical analysis from the perspective of privacy theft, we first explore how to implement the aforementioned ideas. Building upon the findings, we design the Bridge Graph Generator to construct the Shadow-Target Bridge Graph. Then, the subgraph around the link is sampled by the Edge Subgraph Preparation Module. Finally, the Edge Structure Feature Extractor is designed to obtain generalized structure knowledge, which is combined with node similarity to form the features provided to the attack model. Extensive experiments validate the correctness of theoretical analysis and demonstrate that LinkThief still effectively steals links without extra assumptions. Our code is available at https://github.com/octopusStar218/LinkThief-MM2024.

While RAW images are efficient for image editing and perception tasks, their large size can strain camera storage and bandwidth. Reconstruction methods of RAW images from sRGB data typically require additional metadata from the RAW image, which increases camera processing computations. To address this problem, we propose using Prior Meta as a reference to reconstruct the RAW data instead of relying on per-image metadata. Prior metadata is extracted offline from reference RAW images, which are usually part of the training dataset and have similar scenes and light conditions as the target image. With this prior metadata, the camera does not need to provide any extra processing other than the sRGB images, and our model can autonomously find the desired prior information. To achieve this, we design a three-step pipeline. First, we build a pixel searching network that can find the most similar pixels in the reference RAW images as prior information. Then, in the second step, we compress the large-scale reference images to about 0.02% of their original size to reduce the searching cost. Finally, in the last step, we develop a neural network reconstructor to reconstruct the high-fidelity RAW images. Our model achieves comparable, and even better, performance than RAW reconstruction methods based on metadata.

Cross-view consensus representation plays a critical role in hyperspectral images (HSIs) clustering. Recent multi-view contrastive cluster methods utilize contrastive loss to extract contextual consensus representation. However, these methods have a fatal flaw: contrastive learning may treat similar heterogeneous views as positive sample pairs and dissimilar homogeneous views as negative sample pairs. At the same time, the data representation via self-supervised contrastive loss is not specifically designed for clustering. Thus, to tackle this challenge, we propose a novel multi-view clustering method, i.e., Enhanced Multi-View Contrastive Clustering (EMVCC). First, the spatial multi-view is designed to learn the diverse features for contrastive clustering, and the globally relevant information of spectrum-view is extracted by Transformer, enhancing the spatial multi-view differences between neighboring samples. Then, a joint self-supervised loss is designed to constrain the consensus representation from different perspectives to efficiently avoid false negative pairs. Specifically, to preserve the diversity of multi-view information, the features are enhanced by using probabilistic contrastive loss, and the data is projected into a semantic representation space, ensuring that the similar samples in this space are closer in distance. Finally, we design a novel clustering loss that aligns the view feature representation with high confidence pseudo-labels for promoting the network to learn cluster-friendly features. In the training process, the joint self-supervised loss is used to optimize the cross-view features.Abundant experiment studies on numerous benchmarks verify the superiority of EMVCC in comparison to some state-of-the-art clustering methods. The codes are available at https://github.com/YiLiu1999/EMVCC.

Due to its high speed and low latency, DVS is frequently employed in motion deblurring. Ideally, high-quality events would adeptly capture intricate motion information. However, real-world events are generally degraded, thereby introducing significant artifacts into the deblurred results. In response to this challenge, we model the degradation of events and propose RDNet to improve the quality of image deblurring. Specifically, we first analyze the mechanisms underlying degradation and simulate paired events based on that. These paired events are then fed into the first stage of the RDNet for training the restoration model. The events restored in this stage serve as a guide for the second-stage deblurring process. To better assess the deblurring performance of different methods on real-world degraded events, we present a new real-world dataset named DavisMCR. This dataset incorporates events with diverse degradation levels, collected by manipulating environmental brightness and target object contrast. Our experiments are conducted on synthetic datasets (GOPRO), real-world datasets (REBlur), and the proposed dataset (DavisMCR). The results demonstrate that RDNet outperforms classical event denoising methods in event restoration. Furthermore, RDNet exhibits better performance in deblurring tasks compared to state-of-the-art methods. DavisMCR are available at https://github.com/Yeeesir/DVS_RDNet.

Radiology report generation aims to automatically generate clinical descriptions for radiology images, reducing the workload of radiologists. Compared to general image captioning tasks, the subtle differences in medical images and the specialized, complex nature of medical terminology limit the performance of data-driven radiology report generation. Previous research has attempted to leverage prior knowledge, such as organ-disease graphs, to enhance models' abilities to identify specific diseases and generate corresponding medical terminology. However, these methods cover only a limited number of disease types, focusing solely on disease terms mentioned in reports but ignoring their normal or abnormal attributes, which are critical to generating accurate reports. To address this issue, we propose a Divide-and-Conquer approach, named DCG, which separately constructs disease-free and disease-specific nodes within the knowledge graphs. Specifically, we extracted more comprehensive organ-disease entities from reports than previous methods and constructed disease-free and disease-specific nodes by rigorously distinguishing between normal conditions and specific diseases. This enables our model to consciously focus on abnormal information and mitigate the impact of excessively common diseases on report generation. Subsequently, the constructed graph is utilized to enhance the correlation between visual representations and disease terminology, thereby guiding the decoder in report generation. Extensive experiments conducted on benchmark datasets IU-Xray and MIMIC-CXR demonstrate the superiority of our proposed method. Code is available at https://github.com/ecoxial2007/DCG_Enhanced_distilGPT2.

Low-light image enhancement has been researched several years. However, current image restoration methods predominantly focus on recovering images from RGB images, overlooking the potential of incorporating more modalities. With the advancements in personal handheld devices, we can now easily capture images with depth information using devices such as mobile phones. The integration of depth information into image restoration is a research question worthy of exploration. Therefore, in this paper, we propose a multimodal low-light image enhancement task based on depth information and establish a dataset named LED (Low-light Image Enhanced with Depth Map), consisting of 1,365 samples. Each sample in our dataset includes a low-light image, a normal-light image, and the corresponding depth map. Moreover, for the LED dataset, we design a corresponding multimodal method, which can processes the input images and depth map information simultaneously to generate the predicted normal-light images. Experimental results and detailed ablation studies proves the efficiency of our method which exceeds previous single-modal state-of-the arts methods from relevant field.