ACM MM 2024 论文列表

Cloud-based machine learning services offer significant advantages but also introduce the risk of tampering with cloud-deployed deep neural network (DNN) models. Black-box integrity verification (BIV) allows model owners and end-users to determine if a cloud-deployed DNN model has been tampered with by examining only the top-1 label responses. Fingerprinting generates fingerprint samples to query the model, achieving BIV with no impact on the model's accuracy. In this paper, we present BIVBench, the first comprehensive benchmark for BIV of DNN models. BIVBench covers 16 types of model modifications, providing extensive coverage of practical modification scenarios. Our analysis reveals that existing fingerprinting methods, which are typically focused on significant tampering, lack the sensitivity needed to effectively detect subtle yet common and potentially severe modifications. To address this limitation, we propose MiSentry (Model Integrity Sentry), a novel fingerprinting method that leverages meta-learning. MiSentry strategically incorporates a few subtly modified models into the meta-learning model zoo and maximizes the divergence of output predictions between the target model and the modified models in the model zoo to generate highly sensitive, generalizable, and effective fingerprint samples. Extensive evaluations using BIVBench demonstrate that MiSentry outperforms existing state-of-the-art methods overall and significantly surpasses them in detecting subtle modifications. The BIVBench and supplementary materials are available at: https://github.com/CGCL-codes/BIVBench.

Research on continual learning in multi-modal tasks has been receiving increasing attention. However, most existing work overlooks the explicit cross-modal and cross-task interactions. In this paper, we innovatively propose the Low-rank Prompt Interaction (LPI) to address this general problem of multi-modal understanding, which considers both cross-modal and cross-task interactions. Specifically, as for the former, we employ multi-modal correlation modules for corresponding Transformer layers. Considering that the training parameters scale to the number of layers and tasks, we propose low-rank interaction-augmented decomposition to avoid memory explosion while enhancing the cross-modal association through sharing and separating common-specific low-rank factors. In addition, due to the multi-modal semantic differences carried by the low-rank initialization, we adopt hierarchical low-rank contrastive learning to ensure training robustness. As for the latter, we initially employ a visual analysis and identify that different tasks have clear distinctions in proximity. Therefore, we introduce explicit task contrastive constraints in the prompt learning process based on task semantic distances. Experiments on two retrieval tasks show performance improvements with the introduction of a minimal number of parameters, demonstrating the effectiveness of our method. Code is available at https://github.com/Kelvin-ywc/LPI.

Polygonal meshes are widely used to represent complex geometries. However, the increasing complexity of models often leads to large meshes with millions of triangles, raising significant challenges for storage, transmission, and computation. Mesh simplification, a process of reducing the number of triangles in a mesh while preserving its overall shape and important features, has emerged as an indispensable technique to address these challenges. In this work, we focus on the problem of obtaining a visually consistent ultra-low-polygon mesh for complex meshes. Unlike previous methods, we design a robust simplification framework, SimpliGuard, to handle any meshes in the wild. Firstly, a reconstruction module is used to construct a low-polygon mesh with a similar shape but a manifold topology. Then, a texture initialization module is employed to quickly initialize the entire texture map. After that, a differentiable rendering module is utilized to optimize the overall structure and texture details, ensuring high-quality results. For meshes with skeletons, the correctness of motion can be preserved with our designed motion post-processing module. Experimental results demonstrate that SimpliGuard significantly outperforms previous methods and various featured software, including Blender and Simplygon.

Salient Object Detection (SOD) aims to identify and segment the most prominent objects in images. Advanced SOD methods often utilize various Convolutional Neural Networks (CNN) or Transformers for deep feature extraction. However, these methods still deliver low performance and poor generalization in complex cases. Recently, Segment Anything Model (SAM) has been proposed as a visual fundamental model, which gives strong segmentation and generalization capabilities. Nonetheless, SAM requires accurate prompts of target objects, which are unavailable in SOD. Additionally, SAM lacks the utilization of multi-scale and multi-level information, as well as the incorporation of fine-grained details. To address these shortcomings, we propose a Multi-scale and Detail-enhanced SAM (MDSAM) for SOD. Specifically, we first introduce a Lightweight Multi-Scale Adapter (LMSA), which allows SAM to learn multi-scale information with very few trainable parameters. Then, we propose a Multi-Level Fusion Module (MLFM) to comprehensively utilize the multi-level information from the SAM's encoder. Finally, we propose a Detail Enhancement Module (DEM) to incorporate SAM with fine-grained details. Experimental results demonstrate the superior performance of our model on multiple SOD datasets and its strong generalization on other segmentation tasks. The source code is released at https://github.com/BellyBeauty/MDSAM

World models are progressively being employed across diverse fields, extending from basic environment simulation to complex scenario construction. However, existing models are mainly trained on domain-specific states and actions, and confined to single-modality state representations. In this paper, We introduce WorldGPT, a generalist world model built upon Multimodal Large Language Model (MLLM). WorldGPT acquires an understanding of world dynamics through analyzing millions of videos across various domains. To further enhance WorldGPT's capability in specialized scenarios and long-term tasks, we have integrated it with a novel cognitive architecture that combines memory offloading, knowledge retrieval, and context reflection. As for evaluation, we build WorldNet, a multimodal state transition prediction benchmark encompassing varied real-life scenarios. Conducting evaluations on WorldNet directly demonstrates WorldGPT's capability to accurately model state transition patterns, affirming its effectiveness in understanding and predicting the dynamics of complex scenarios. We further explore WorldGPT's emerging potential in serving as a world simulator, helping multimodal agents generalize to unfamiliar domains through efficiently synthesising multimodal instruction instances which are proved to be as reliable as authentic data for fine-tuning purposes. The code and dataset are available on the https://github.com/DCDmllm/WorldGPT

Recently, video diffusion models (VDMs) have garnered significant attention due to their notable advancements in generating coherent and realistic video content. However, processing multiple frame features concurrently, coupled with the considerable model size, results in high latency and extensive memory consumption, hindering their broader application. Post-training quantization (PTQ) is an effective technique to reduce memory footprint and improve computational efficiency. Unlike image diffusion, we observe that the temporal features, which are integrated into all frame features, exhibit pronounced skewness. Furthermore, we investigate significant inter-channel disparities and asymmetries in the activation of video diffusion models, resulting in low coverage of quantization levels by individual channels and increasing the challenge of quantization. To address these issues, we introduce the first PTQ strategy tailored for video diffusion models, dubbed QVD. Specifically, we propose the High Temporal Discriminability Quantization (HTDQ) method, designed for temporal features, which retains the high discriminability of quantized features, providing precise temporal guidance for all video frames. In addition, we present the Scattered Channel Range Integration (SCRI) method which aims to improve the coverage of quantization levels across individual channels. Experimental validations across various models, datasets, and bit-width settings demonstrate the effectiveness of our QVD in terms of diverse metrics. In particular, we achieve near-lossless performance degradation on W8A8, outperforming the current methods by 205.12 in FVD.

3D Referring Expression Segmentation (3D-RES) is dedicated to segmenting a specific instance within a 3D space based on a natural language description. However, current approaches are limited to segmenting a single target, restricting the versatility of the task. To overcome this limitation, we introduce Generalized 3D Referring Expression Segmentation (3D-GRES), which extends the capability to segment any number of instances based on natural language instructions. In addressing this broader task, we propose the Multi-Query Decoupled Interaction Network (MDIN), designed to break down multi-object segmentation tasks into simpler, individual segmentations. MDIN comprises two fundamental components: Text-driven Sparse Queries (TSQ) and Multi-object Decoupling Optimization (MDO). TSQ generates sparse point cloud features distributed over key targets as the initialization for queries. Meanwhile, MDO is tasked with assigning each target in multi-object scenarios to different queries while maintaining their semantic consistency. To adapt to this new task, we build a new dataset, namely Multi3DRes. Our comprehensive evaluations on this dataset demonstrate substantial enhancements over existing models, thus charting a new path for intricate multi-object 3D scene comprehension. The benchmark and code are available at https://github.com/sosppxo/MDIN.

In recent years, the Few-Shot Fine-Grained Image Classification (FS-FGIC) problem has gained widespread attention. A number of effective methods have been proposed that focus on extracting discriminative information within high-level features in a single episode/task. However, this is insufficient for addressing the cross-task challenges of FS-FGIC, which is represented in two aspects. On the one hand, from the perspective of the Fine-Grained Image Classification (FGIC) task, there is a need to supplement the model with mid-level features containing rich fine-grained information. On the other hand, from the perspective of the Few-Shot Learning (FSL) task, explicit modeling of cross-task general knowledge is required. In this paper, we propose a novel Bi-directional Task-Guided Network (BTG-Net) to tackle these issues. Specifically, from the FGIC task perspective, we design the Semantic-Guided Noise Filtering (SGNF) module to filter noise on mid-level features rich in detailed information. Further, from the FSL task perspective, the General Knowledge Prompt Modeling (GKPM) module is proposed to retain the cross-task general knowledge by utilizing the prompting mechanism, thereby enhancing the model's generalization performance on novel classes. We have conducted extensive experiments on five fine-grained benchmark datasets, and the results demonstrate that BTG-Net outperforms state-of-the-art methods comprehensively.

With the rapid development of high-bit-depth display devices, bit-depth expansion (BDE) algorithms that extend low-bit-depth images to high-bit-depth images have received increasing attention. Due to the sensitivity of bit-depth distortions to tiny numerical changes in the least significant bits, the nuanced degradation differences in the training process may lead to varying degradation data distributions, causing the trained models to overfit specific types of degradations. This paper focuses on the problem of blind video BDE, proposing a degradation prediction and embedding framework, and designing a video BDE network based on a recurrent structure and dual-frame alignment fusion. Experimental results demonstrate that the proposed model can outperform some state-of-the-art (SOTA) models in terms of banding artifact removal and color correction, avoiding overfitting to specific degradations and obtaining better generalization ability across multiple datasets. https://github.com/duanpanjun/BVBDE

Deep Video Quality Assessment (VQA) methods have shown impressive high-performance capabilities. Notably, no-reference (NR) VQA methods play a vital role in situations where obtaining reference videos is restricted or not feasible. Nevertheless, as more streaming videos are being created in ultra-high definition (e.g., 4K) to enrich viewers' experiences, the current deep VQA methods face unacceptable computational costs. Furthermore, the resizing, cropping, and local sampling techniques employed in these methods can compromise the details and content of original 4K videos, thereby negatively impacting quality assessment. In this paper, we propose a highly efficient and novel NR 4K VQA technology. Specifically, first, a novel data sampling and training strategy is proposed to tackle the problem of excessive resolution. This strategy allows the VQA Swin Transformer-based model to effectively train and make inferences using the full data of 4K videos on standard consumer-grade GPUs without compromising content or details. Second, a weighting and scoring scheme is developed to mimic the human subjective perception mode, which is achieved by considering the distinct impact of each sub-region within a 4K frame on the overall perception. Third, we incorporate the frequency domain information of video frames to better capture the details that affect video quality, consequently further improving the model's generalizability. To our knowledge, this is the first technology for the NR 4K VQA task. Thorough empirical studies demonstrate it not only significantly outperforms existing methods on a specialized 4K VQA dataset but also achieves state-of-the-art performance across multiple open-source NR video quality datasets.

Spatial transcriptomics provides revolutionary insights into cellular interactions and disease development mechanisms by combining high-throughput gene sequencing and spatially resolved imaging technologies to analyze genes naturally associated with spatially variable tissue genes. However, existing methods typically map aggregated multi-view features into a unified representation, ignoring the heterogeneity and view independence of genes and spatial information. To this end, we construct a heterogeneous Graph guided Contrastive Learning (stGCL) for aggregating spatial transcriptomics data. The method is guided by the inherent heterogeneity of cellular molecules by dynamically coordinating triple-level node attributes through comparative learning loss distributed across view domains, thus maintaining view independence during the aggregation process. In addition, we introduce a cross-view hierarchical feature alignment module employing a parallel approach to decouple spatial and genetic views on molecular structures while aggregating multi-view features according to information theory, thereby enhancing the integrity of inter- and intra-views. Rigorous experiments demonstrate that stGCL outperforms existing methods in various tasks and related downstream applications.

Cross-lingual cross-modal retrieval (CCR) aims to retrieve visually relevant content based on non-English queries, without relying on human-labeled cross-modal data pairs during training. One popular approach involves utilizing machine translation (MT) to create pseudo-parallel data pairs, establishing correspondence between visual and non-English textual data. However, aligning their representations poses challenges due to the significant semantic gap between vision and text, as well as the lower quality of non-English representations caused by pre-trained encoders and data noise. To overcome these challenges, we propose LECCR, a novel solution that incorporates the multi-modal large language model (MLLM) to improve the alignment between visual and non-English representations. Specifically, we first employ MLLM to generate detailed visual content descriptions and aggregate them into multi-view semantic slots that encapsulate different semantics. Then, we take these semantic slots as internal features and leverage them to interact with the visual features. By doing so, we enhance the semantic information within the visual features, narrowing the semantic gap between modalities and generating local visual semantics for subsequent multi-level matching. Additionally, to further enhance the alignment between visual and non-English features, we introduce softened matching under English guidance. This approach provides more comprehensive and reliable inter-modal correspondences between visual and non-English features. Extensive experiments on four CCR benchmarks, i.e., Multi30K, MSCOCO, VATEX, and MSR-VTT-CN, demonstrate the effectiveness of our proposed method. Code: https://github.com/LiJiaBei-7/leccr.

This paper presents the first study to explore the potential of parameter quantization for multimodal large language models to alleviate the significant resource constraint encountered during vision-language instruction tuning. We introduce a Quantization-aware Scale LeArning method based on multimodal Warmup, termed QSLAW. This method is grounded in two key innovations: (1) The learning of group-wise scale factors for quantized LLM weights to mitigate the quantization error arising from activation outliers and achieve more effective vision-language instruction tuning; (2) The implementation of a multimodal warmup that progressively integrates linguistic and multimodal training samples, thereby preventing overfitting of the quantized model to multimodal data while ensuring stable adaptation of multimodal large language models to downstream vision-language tasks. Extensive experiments demonstrate that models quantized by QSLAW perform on par with, or even surpass, their full-precision counterparts, while facilitating up to 1.4 times reduction in VL tuning time and GPU consumption.

Latent diffusion model has demonstrated impressive efficacy in image generation and editing tasks. Recently, it has also promoted the advancement of image harmonization. However, methods involving latent diffusion model all face a common challenge: the severe image distortion introduced by the VAE component, while image harmonization is a low-level image processing task that relies on pixel-level evaluation metrics. In this paper, we propose Harmony-VAE, leveraging the input of the harmonization task itself to enhance the quality of decoded images. The input involving composite image contains the precise pixel level information, which can complement the correct foreground appearance and color information contained in denoised latents. Meanwhile, the inherent generative nature of diffusion models makes it naturally adapt to inverse image harmonization, i.e. generating synthetic composite images based on real images and foreground masks. We train an inverse harmonization diffusion model to perform data augmentation on two subsets of iHarmony4 and construct a new human harmonization dataset with prominent foreground objects. Extensive experiments demonstrate the effectiveness of our proposed Harmony-VAE and inverse harmonization model. Code and pretrained models are available at https://github.com/nicecv/DiffHarmony.

Sound Effect (SFX) generation, primarily aims to automatically produce sound waves for sounding visual objects in images or videos. Rather than learning an automatic solution to this task, we aim to propose a much broader system, AutoSFX, designed to automate sound design for videos in a more efficient and applicable manner. AutoSFX capitalizes on this concept by aggregating multimodal representations by cross-attention and leverages a diffusion model to generate sound with visual information embedded. AutoSFX also optimizes the generated sounds to render the entire soundtrack for the input video, leading to a more immersive and engaging multimedia experience by performing seamless transitions between sound clips and harmoniously mixing sounds playing simultaneously. We have developed a user-friendly interface for AutoSFX enabling users to interactively engage in the SFX generation for their videos with particular needs. To validate the capability of our vision-to-sound generation, we conducted comprehensive experiments and analyses using the widely recognized VEGAS and VGGSound test sets, yielding promising results. We also conducted a user study to evaluate the performance of the optimized soundtrack and the usability of the interface. Overall, the results revealed that our AutoSFX provides a viable sound landscape solution for making attractive videos.

While significant progress has been made in multi-modal learning driven by large-scale image-text datasets, there is still a noticeable gap in the availability of such datasets within the facial domain. To facilitate and advance the field of facial representation learning, we present FLIP-80M, a large-scale visual-linguistic dataset comprising over 80 million face images paired with text descriptions. FLIP-80M is constructed by leveraging the large openly available image-text-pair dataset LAION-5B and a mixed-method approach to filter face-related pairs from both visual and linguistic perspectives. Our curation process involves face detection, face caption classification, text de-noising, and synthesis-based image augmentation. As a result, FLIP-80M stands as the largest face-text dataset to date. To evaluate the potential of our dataset, we fine-tune the CLIP model using the proposed FLIP-80M, to create FLIP (Facial Language-Image Pretraining) and assess its representation capabilities across various downstream tasks. Our experiments demonstrate that our FLIP model achieves state-of-the-art results in a range of face analysis tasks, including face parsing, face alignment, and face attribute classification. The dataset and models are available at https://github.com/ydli-ai/FLIP.