多模态大模型

Semantic segmentation is a crucial prerequisite in clinical applications and computer-aided diagnosis. With the development of deep neural networks, biomedical image segmentation has achieved remarkable success. Encoder-decoder architectures that integrate convolutions and transformers are gaining attention for their potential to capture both global and local features. However, current designs face the contradiction that these two features cannot be continuously transmitted. In addition, some models lack a unified and standardized evaluation benchmark, leading to significant discrepancies in the experimental setup. In this study, we review and summarize these architectures and analyze their contradictions in design. We modify UNet and propose WNet to combine transformers and convolutions, addressing the transmission issue effectively. WNet captures long-range dependencies and local details simultaneously while ensuring their continuous transmission and multi-scale fusion. We integrate WNet into the nnUNet framework for unified benchmarking. Our model achieves state-of-the-art performance in biomedical image segmentation. Extensive experiments demonstrate their effectiveness on four 2D datasets (DRIVE, ISIC-2017, Kvasir-SEG, and CREMI) and four 3D datasets (Parse2022, AMOS22, BTCV, and ImageCAS). The code is available at https://github.com/Yanfeng-Zhou/nnWNet.

We introduce Generalizable 3D-Language Feature Fields (g3D-LF), a 3D representation model pre-trained on large-scale 3D-language dataset for embodied tasks. Our g3D-LF processes posed RGB-D images from agents to encode feature fields for: 1) Novel view representation predictions from any position in the 3D scene; 2) Generations of BEV maps centered on the agent; 3) Querying targets using multi-granularity language within the above-mentioned representations. Our representation can be generalized to unseen environments, enabling real-time construction and dynamic updates. By volume rendering latent features along sampled rays and integrating semantic and spatial relationships through multiscale encoders, our g3D-LF produces representations at different scales and perspectives, aligned with multi-granularity language, via multi-level contrastive learning. Furthermore, we prepare a large-scale 3D-language dataset to align the representations of the feature fields with language. Extensive experiments on Vision-and-Language Navigation under both Panorama and Monocular settings, Zero-shot Object Navigation, and Situated Question Answering tasks highlight the significant advantages and effectiveness of our g3D-LF for embodied tasks. Our source code and dataset will be made open-source upon paper acceptance.

Co-training has achieved significant success in the field of semi-supervised learning; however, the *homogenization phenomenon*, which arises from multiple models tending towards similar decision boundaries, remains inadequately addressed. To tackle this issue, we propose a novel algorithm called **β-FFT** from the perspectives of data diversity and training structure.First, from the perspective of data diversity, we introduce a nonlinear interpolation method based on the **Fast Fourier Transform (FFT)**. This method generates more diverse training samples by swapping low-frequency components between pairs of images, thereby enhancing the model's generalization capability. Second, from the structural perspective, we propose a differentiated training strategy to alleviate the homogenization issue in co-training. In this strategy, we apply additional training with labeled data to one model in the co-training framework, while employing linear interpolation based on the **Beta (β)** distribution for the unlabeled data as a regularization term for the additional training. This approach enables us to effectively utilize the limited labeled data while simultaneously improving the model's performance on unlabeled data, ultimately enhancing the overall performance of the system.Experimental results demonstrate that **β-FFT** outperforms current state-of-the-art (SOTA) methods on three public medical image datasets.

Recent advancements in 3D Large Multi-modal Models (3D-LMMs) have driven significant progress in 3D question answering. However, recent multi-frame Vision-Language Models (VLMs) demonstrate superior performance compared to 3D-LMMs on 3D question answering tasks, largely due to the greater scale and diversity of available 2D image data in contrast to the more limited 3D data. Multi-frame VLMs, although achieving superior performance, suffer from the difficulty of retaining all the detailed visual information in the 3D scene while limiting the number of visual tokens. Common methods such as token pooling, reduce visual token usage but often lead to information loss, impairing the model's ability to preserve visual details essential for 3D question answering tasks. To address this, we propose voxel-based Dynamic Token Compression (DTC), which combines 3D spatial priors and visual semantics to achieve over 90% reduction in visual tokens usage for current multi-frame VLMs. Our method maintains performance comparable to state-of-the-art models on 3D question answering benchmarks including OpenEQA and ScanQA, demonstrating its effectiveness.

Zero-shot 4D segmentation and recognition of arbitrary objects in Lidar is crucial for embodied navigation, with applications ranging from streaming perception to semantic mapping and localization. However, the primary challenge in advancing research and developing generalized, versatile methods for spatio-temporal scene understanding in Lidar lies in the scarcity of datasets that provide the necessary diversity and scale of annotations. To overcome these challenges, we propose SAL-4D (Segment Anything in Lidar--4D), a method that utilizes multi-modal robotic sensor setups as a bridge to distill recent developments in Video Object Segmentation (VOS) in conjunction with off-the-shelf Vision-Language foundation models to Lidar. We utilize VOS models to pseudo-label tracklets in short video sequences, annotate these tracklets with sequence-level CLIP tokens, and lift them to the 4D Lidar space using calibrated multi-modal sensory setups to distill them to our SAL-4D model. Due to temporal consistent predictions, we outperform prior art in 3D Zero-Shot Lidar Panoptic Segmentation (LPS) over 5 PQ, and unlock Zero-Shot 4D-LPS.

Visual grounding seeks to localize the image region corresponding to a free-form text description. Recently, the strong multimodal capabilities of Large Vision-Language Models (LVLMs) have driven substantial improvements in visual grounding, though they inevitably require fine-tuning and additional model components to explicitly generate bounding boxes or segmentation masks. However, we discover that a few attention heads in frozen LVLMs demonstrate strong visual grounding capabilities. We refer to these heads, which consistently capture object locations related to text semantics, as localization heads. Using localization heads, we introduce a straightforward and effective training-free visual grounding framework that utilizes text-to-image attention maps from localization heads to identify the target objects. Surprisingly, only three out of thousands of attention heads are sufficient to achieve competitive localization performance compared to existing LVLM-based visual grounding methods that require fine-tuning. Our findings suggest that LVLMs can innately ground objects based on a deep comprehension of the text-image relationship, as they implicitly focus on relevant image regions to generate informative text outputs.

Large Multimodal Models (e.g., GPT-4, Gemini, Chameleon) have evolved into powerful tools with millions of users. However, they remain generic models and lack personalized knowledge of specific user concepts. Previous work has explored personalization for text generation, yet it remains unclear how these methods can be adapted to new modalities, such as image generation. In this paper, we introduce Yo'Chameleon, the first attempt to study personalization for large multimodal models. Given 3-5 images of a particular concept, Yo'Chameleon leverages soft-prompt tuning to embed subject-specific information to (i) answer questions about the subject and (ii) recreate pixel-level details to produce images of the subject in new contexts. Yo'Chameleon is trained with (i) a self-prompting optimization mechanism to balance performance across multiple modalities, and (ii) a "soft-positive" image generation approach to enhance image quality in a few-shot setting. Our qualitative and quantitative analyses reveal that Yo'Chameleon can learn concepts more efficiently using fewer tokens and effectively encode visual attributes, outperforming prompting baselines.

The astonishing breakthrough of multimodal large language models (MLLMs) has necessitated new benchmarks to quantitatively assess their capabilities, reveal their limitations, and indicate future research directions. However, this is challenging in the context of remote sensing (RS), since the imagery features ultra-high resolution that incorporates extremely complex semantic relationships. Existing benchmarks usually adopt notably smaller image sizes than real-world RS scenarios, suffer from limited annotation quality, and consider insufficient dimensions of evaluation. To address these issues, we present XLRS-Bench: a comprehensive benchmark for evaluating the perception and reasoning capabilities of MLLMs in ultra-high-resolution RS scenarios. XLRS-Bench boasts the largest average image size (8500x8500) observed thus far, with all evaluation samples meticulously annotated manually, assisted by a novel semi-automatic captioner on ultra-high-resolution RS images. On top of the XLRS-Bench, 16 sub-tasks are defined to evaluate MLLMs' 10 kinds of perceptual capabilities and 6 kinds of reasoning capabilities, with a primary emphasis on advanced cognitive processes that facilitate real-world decision-making and the capture of spatiotemporal changes. The results of both general and RS-focused MLLMs on XLRS-Bench indicate that further efforts are needed for real-world RS applications. We will open source XLRS-Bench to support further research of developing more powerful MLLMs for RS.

Vision-Language Models (VLMs) excel in integrating visual and textual information for vision-centric tasks, but their handling of inconsistencies between modalities is underexplored. We investigate VLMs' modality preferences when faced with visual data and varied textual inputs in vision-centered settings.By introducing textual variations to four vision-centric tasks and evaluating ten Vision-Language Models (VLMs), we discover a "blind faith in text" phenomenon: VLMs disproportionately trust textual data over visual data when inconsistencies arise, leading to significant performance drops under corrupted text and raising safety concerns.We analyze factors influencing this text bias, including instruction prompts, language model size, text relevance, token order, and the interplay between visual and textual certainty. While certain factors, such as scaling up the language model size, slightly mitigate text bias, others like token order can exacerbate it due to positional biases inherited from language models. To address this issue, we explore supervised fine-tuning with text augmentation and demonstrate its effectiveness in reducing text bias. Additionally, we provide a theoretical analysis suggesting that the blind faith in text phenomenon may stem from an imbalance of pure text and multi-modal data during training.Our findings highlight the need for balanced training and careful consideration of modality interactions in VLMs to enhance their robustness and reliability in handling multi-modal data inconsistencies.

Vision-Language Models (VLMs) have recently demonstrated remarkable capabilities in comprehending complex visual content. However, the mechanisms underlying how VLMs process visual information remain largely unexplored. In this paper, we conduct a thorough empirical analysis, focusing on the attention modules across layers, by which we reveal several key insights about how these models process visual data: (i) the internal representation of the query tokens (e.g., representations of "describe the image"), is utilized by the model to store global image information; we demonstrate that the model generates surprisingly descriptive responses solely from these tokens, without direct access to image tokens. (ii) Cross-modal information flow is predominantly influenced by the middle layers (approximately 25% of all layers), while early and late layers contribute only marginally. (iii) Fine-grained visual attributes and object details are directly extracted from image tokens in a spatially localized manner, i.e., the generated tokens associated with a specific object or attribute attend strongly to their corresponding regions in the image. We propose novel quantitative evaluation to validate our observations, leveraging real-world complex visual scenes. Finally, we demonstrate the potential of our findings in facilitating efficient visual processing in state-of-the-art VLMs.

Recent breakthroughs in Multimodal Large Language Models (MLLMs) have gained significant recognition within the deep learning community, where the fusion of the Video Foundation Models (VFMs) and Large Language Models(LLMs) has proven instrumental in constructing robust video understanding systems, effectively surmounting constraints associated with predefined visual tasks. These sophisticated MLLMs exhibit remarkable proficiency in comprehending videos, swiftly attaining unprecedented performance levels across diverse benchmarks. However, their operation demands substantial memory and computational resources, underscoring the continued importance of traditional models in video comprehension tasks. In this paper, we introduce a novel learning paradigm termed MLLM4WTAL. This paradigm harnesses the potential of MLLM to offer temporal action key semantics and complete semantic textual cues for conventional Weakly-supervised Temporal Action Localization (WTAL) methods. MLLM4WTAL facilitates the enhancement of WTAL by leveraging MLLM guidance. It achieves this by integrating two distinct modules: Key Semantic Matching (KSM) and Complete Semantic Reconstruction (CSR). These modules work in tandem to effectively address prevalent issues like incomplete and over-complete outcomes common in WTAL methods. Rigorous experiments are conducted to validate the efficacy of our proposed approach in augmenting the performance of various heterogeneous WTAL models.

Existing multimodal generative models fall short as qualified design copilots, as they often struggle to generate imaginative outputs once instructions are less detailed or lack the ability to maintain consistency with the provided references. In this work, we introduce WeGen, a model that unifies multimodal generation and understanding, and promotes their interplay in iterative generation. It can generate diverse results with high creativity for less detailed instructions. And it can progressively refine prior generation results or integrating specific contents from references following the instructions in its chat with users. During this process, it is capable of preserving consistency in the parts that the user is already satisfied with. To this end, we curate a large-scale dataset, extracted from Internet videos, containing rich object dynamics and auto-labeled dynamics descriptions by advanced foundation models to date. These two information are interleaved into a single sequence to enable WeGen to learn consistency-aware generation where the specified dynamics are generated while the consistency of unspecified content is preserved aligned with instructions. Besides, we introduce a prompt self-rewriting mechanism to enhance generation diversity. Extensive experiments demonstrate the effectiveness of unifying multimodal understanding and generation in WeGen and show it achieves state-of-the-art performance across various visual generation benchmarks. These also demonstrate the potential of WeGen as a user-friendly design copilot as desired.

The proliferation of digital images on the Internet has provided unprecedented convenience, but also poses significant risks of malicious theft and misuse. Digital watermarking has long been researched as an effective tool for copyright protection. However, it often falls short when addressing partial image theft, a common yet little-researched issue in practical applications. Most existing schemes typically require the entire image as input to extract watermarks. However, in practice, malicious users often steal only a portion of the image to create new content. The stolen portion can have arbitrary shape or content, being fused with a new background and may have undergone geometric transformations, making it challenging for current methods to extract correctly. To address the issues above, we propose WOFA (Watermarking One for All), a robust watermarking scheme against partial image theft. First of all, we define the entire process of partial image theft and construct a dataset accordingly. To gain robustness against partial image theft, we then design a comprehensive distortion layer that incorporates the process of partial image theft and several common distortions in channel. For easier network convergence, we employ a multi-level network structure on the basis of the commonly used embedder-distortion layer-extractor architecture and adopt a progressive training strategy. Abundant experiments demonstrate that our superior performance in the scenario of partial image theft, offering a more reliable solution for protecting digital images against unauthorized use in practical use.

Whole-Slide Images (WSIs) have revolutionized medical analysis by presenting high-resolution images of the whole tissue slide. Despite avoiding the physical storage of the slides, WSIs require considerable data volume, which makes the storage and maintenance of WSI records costly and unsustainable. To this end, this work presents the first investigation of lossless compression of WSI images. Interestingly, we find that most existing compression methods fail to compress the WSI images effectively. Furthermore, our analysis reveals that the failure of existing compressors is mainly due to information irregularity in WSI images. To resolve this issue, we develop a simple yet effective lossless compressor called WISE, specifically designed for WSI images. WISE employs a hierarchical encoding strategy to extract effective bits, reducing the entropy of the image and then adopting a dictionary-based method to handle the irregular frequency patterns. Through extensive experiments, we show that WISE can effectively compress the gigapixel WSI images to 36 times on average and up to 136 times.

Capitalizing on the talent of self-attention in capturing non-local features, Transformer architectures have exhibited remarkable performance in single hyperspectral image restoration. For hyperspectral images, each pixel is located in the hyperspectral image cubes with a large spectral dimension and two spatial dimensions. Although uni-dimensional self-attention, like channel self-attention or spatial self-attention, builds long-range dependencies in spectral or spatial dimensions, they lack more comprehensive interactions across dimensions. To tackle the above drawback, we propose a VolFormer, a volumetric self-attention embedded Transformer network for single hyperspectral image restoration. Specifically, we propose volumetric self-attention (VolSA), which extends the interaction from 2D flat to 3D cube. VolSA can simultaneously model token interaction in the 3D cube, mining the potential correlations between the hyperspectral image cube. An attention decomposition form is proposed to reduce the computational burden of modeling volumetric information. In practical terms, VolSA adapts double similarity matrixes in spatial and channel dimensions to implicitly model 3D context information while transforming the complexity from cubic to quadratic. Additionally, we introduce the explicit spectral location prior to enhance the proposed self-attention. This property allows the target token to perceive global spectral information while simultaneously assigning different levels of attention to tokens at varying wavelength bands. Extensive experiments demonstrate that VolFormer achieves record-high performance on hyperspectral image super-resolution, denoise and classification benchmarks. Particularly, VolSA is portable and achieves inspiring results in hyperspectral classification. The source code is available at https://github.com/yudadabing/VolFormer.

Vision-Language Models (VLMs) have achieved remarkable success in various multi-modal tasks, but they are often bottlenecked by the limited context window and high computational cost of processing high-resolution image inputs and videos. Vision compression can alleviate this problem by reducing the vision token count. Previous approaches compress vision tokens with external modules and force LLMs to understand the compressed ones, leading to visual information loss. However, the LLMs' understanding paradigm of vision tokens is not fully utilised in the compression learning process. We propose VoCo-LLaMA, the first approach to compress vision tokens using LLMs. By introducing Vision Compression tokens during the vision instruction tuning phase and leveraging attention distillation, our method distill how LLMs comprehend vision tokens into their processing of VoCo tokens. VoCo-LLaMA facilitates effective vision compression and improves the computational efficiency during the inference stage. Specifically, our method can achieve a 576x compression rate while maintaining 83.7% performance. Furthermore, through continuous training using time-series compressed token sequences of video frames, VoCo-LLaMA demonstrates the ability to understand temporal correlations, outperforming previous methods on popular video question-answering benchmarks. Our approach presents a promising way to unlock the full potential of VLMs' contextual window, enabling more scalable multi-modal applications. The project page can be accessed via https://yxxxb.github.io/VoCo-LLaMA-page.

Contrastively-trained Vision-Language Models (VLMs) like CLIP have become the de facto approach for discriminative vision-language representation learning. However, these models have limited language understanding, often exhibiting a "bag of words" behavior. At the same time, Large Vision-Language Models (LVLMs), which combine vision encoders with LLMs, have been shown to be capable of detailed vision-language reasoning, yet their autoregressive nature renders them less suitable for discriminative tasks. In this work, we propose to combine "the best of both worlds": a new training approach for discriminative fine-tuning of LVLMs that results in strong discriminative and compositional capabilities. Essentially, our approach converts a generative LVLM into a discriminative one, unlocking its capability for powerful image-text discrimination combined with enhanced language understanding. Our contributions include (1) A carefully designed training/optimization framework that utilizes image-text pairs of variable length and granularity for training the model with both contrastive and next-token prediction losses. This is accompanied by ablation studies that justify the necessity of our framework's components. (2) A parameter-efficient adaptation method using a combination of soft prompting and LoRA adapters. (3) Significant improvements over state-of-the-art CLIP-like models of similar size, including standard image-text retrieval benchmarks and notable gains in compositionality.

Image restoration tasks like deblurring, denoising, and dehazing usually need distinct models for each degradation type, restricting their generalization in real-world scenarios with mixed or unknown degradations. In this work, we propose Defusion, a novel all-in-one image restoration framework that utilizes visual instruction-guided degradation diffusion. Unlike existing methods that rely on task-specific models or ambiguous text-based priors, Defusion constructs explicit visual instructions that align with the visual degradation patterns. These instructions are grounded by applying degradations to standardized visual elements, capturing intrinsic degradation features while agnostic to image semantics. Defusion then uses these visual instructions to guide a diffusion-based model that operates directly in the degradation space, where it reconstructs high-quality images by denoising the degradation effects with enhanced stability and generalizability. Comprehensive experiments demonstrate that Defusion outperforms state-of-the-art methods across diverse image restoration tasks, including complex and real-world degradations.

A key challenge for controllable image editing is that visual attributes with semantic meanings are not always independent, resulting in spurious correlations in model training. However, most existing methods ignore such issues, leading to biased causal visual representation learning and unintended changes to unrelated regions or attributes in the edited images. To bridge this gap, we propose a diffusion-based causal visual representation learning framework called CIDiffuser to capture causal representations of visual attributes based on structural causal models to address the spurious correlation. Specifically, we first decompose the image representation into a high-level semantic representation for core attributes of the image and a low-level stochastic representation for other random or less structured aspects, with the former extracted by a semantic encoder and the latter derived via a stochastic encoder. We then introduce a causal effect learning module to capture the direct causal effect, that is, the difference of potential outcomes before and after intervening on the visual attributes. In addition, a diffusion-based learning strategy is designed to optimize the representation learning process. Empirical evaluations on two benchmark datasets demonstrate that our approach significantly outperforms state-of-the-art methods, enabling highly controllable image editing by modifying learned visual representations.