Current knowledge distillation (KD) methods for semantic segmentation focus on guiding the student to imitate the teacher's knowledge within homogeneous architectures. However, these methods overlook the diverse knowledge contained in architectures with different inductive biases, which is crucial for enabling the student to acquire a more precise and comprehensive understanding of the data during distillation. To this end, we propose for the first time a generic knowledge distillation method for semantic segmentation from a heterogeneous perspective, named HeteroAKD. Due to the substantial disparities between heterogeneous architectures, such as CNN and Transformer, directly transferring cross-architecture knowledge presents significant challenges. To eliminate the influence of architecture-specific information, the intermediate features of both the teacher and student are skillfully projected into an aligned logits space. Furthermore, to utilize diverse knowledge from heterogeneous architectures and deliver customized knowledge required by the student, a teacher-student knowledge mixing mechanism (KMM) and a teacher-student knowledge evaluation mechanism (KEM) are introduced. These mechanisms are performed by assessing the reliability and its discrepancy between heterogeneous teacher-student knowledge. Extensive experiments conducted on three main-stream benchmarks using various teacher-student pairs demonstrate that our HeteroAKD framework outperforms state-of-the-art KD methods in facilitating distillation between heterogeneous architectures.
论文检索
输入标题、作者或关键词，从 11,272 篇学术成果中精准定位
Low-light image enhancement (LLIE) aims to improve visibility and signal-to-noise ratio in images captured under poor lighting conditions. While deep learning has shown promise in this domain, current approaches require extensive paired training data, limiting their practical utility. We present a novel framework that reformulates low-light image enhancement as a zero-shot inference problem using pre-trained latent diffusion models (LDMs), eliminating the need for task-specific training data. Our key insight is that the rich natural image priors encoded in LDMs can be leveraged to recover well-lit images through a carefully designed optimization process. To address the ill-posed nature of low-light degradation and the complexity of latent space optimization, our framework introduces an exposure-aware degradation module that adaptively models illumination variations and a principled latent regularization scheme with adaptive guidance that ensures both enhancement quality and natural image statistics. Experimental results demonstrate that our framework outperforms existing zero-shot methods across diverse real-world scenarios.
LiDAR-based 3D object detection is crucial for autonomous driving. However, due to the quality deterioration of LiDAR point clouds, it suffers from performance degradation in adverse weather conditions. Fusing LiDAR with the weatherrobust 4D radar sensor is expected to solve this problem; however, it faces challenges of significant differences in terms of data quality and the degree of degradation in adverse weather. To address these issues, we introduce L4DR, a weather-robust 3D object detection method that effectively achieves LiDAR and 4D Radar fusion. Our L4DR proposes Multi-Modal Encoding (MME) and Foreground-Aware Denoising (FAD) modules to reconcile sensor gaps, which is the first exploration of the complementarity of early fusion between LiDAR and 4D radar. Additionally, we design an Inter-Modal and IntraModal ({IM}2) parallel feature extraction backbone coupled with a Multi-Scale Gated Fusion (MSGF) module to counteract the varying degrees of sensor degradation under adverse weather conditions. Experimental evaluation on a VoD dataset with simulated fog proves that L4DR is more adaptable to changing weather conditions. It delivers a significant performance increase under different fog levels, improving the 3D mAP by up to 20.0% over the traditional LiDAR-only approach. Moreover, the results on the K-Radar dataset validate the consistent performance improvement of L4DR in realworld adverse weather conditions.
Medical MLLM Is Vulnerable: Cross-Modality Jailbreak and Mismatched Attacks on Medical Multimodal Large Language Models
PDF ↗Security concerns related to Large Language Models (LLMs) have been extensively explored; however, the safety implications for Multimodal Large Language Models (MLLMs), particularly in medical contexts (MedMLLMs), remain inadequately addressed. This paper investigates the security vulnerabilities of MedMLLMs, focusing on their deployment in clinical environments where the accuracy and relevance of question-and-answer interactions are crucial for addressing complex medical challenges. We introduce and redefine two attack types: mismatched malicious attack (2M-attack) and optimized mismatched malicious attack (O2M-attack), by integrating existing clinical data with atypical natural phenomena. Using the comprehensive 3MAD dataset that we developed, which spans a diverse range of medical imaging modalities and adverse medical scenarios, we performed an in-depth analysis and proposed the MCM optimization method. This approach significantly improves the attack success rate against MedMLLMs. Our evaluations, which include white-box attacks on LLaVA-Med and transfer (black-box) attacks on four other SOTA models, reveal that even MedMLLMs designed with advanced security mechanisms remain vulnerable to breaches. This study highlights the critical need for robust security measures to enhance the safety and reliability of open-source MedMLLMs, especially in light of the potential impact of jailbreak attacks and other malicious exploits in clinical applications. Warning: Medical jailbreaking may generate content that includes unverified diagnoses and treatment recommendations. Always consult professional medical advice.
The discriminative feature is crucial for point cloud registration. Recent methods improve the feature discriminative by distinguishing between non-overlapping and overlapping region points. However, they still face challenges in distinguishing the ambiguous structures in the overlapping regions. Therefore, the ambiguous features they extracted resulted in a significant number of outlier matches from overlapping regions. To solve this problem, we propose a prior-guided SMoE-based registration method to improve the feature distinctiveness by dispatching the potential correspondences to the same experts. Specifically, we propose a prior-guided SMoE module by fusing prior overlap and potential correspondence embeddings for routing, assigning tokens to the most suitable experts for processing. In addition, we propose a registration framework by a specific combination of Transformer layer and prior-guided SMoE module. The proposed method not only pays attention to the importance of locating the overlapping areas of point clouds, but also commits to finding more accurate correspondences in overlapping areas. Our extensive experiments demonstrate the effectiveness of our method, achieving state-of-the-art registration recall (95.7%/79.3%) on the 3DMatch/3DLoMatch benchmark. Moreover, we also test the performance on ModelNet40 and demonstrate excellent performance.
In recent years, Multimodal Large Language Models (MLLM) have achieved notable advancements, demonstrating the feasibility of developing an intelligent biomedical assistant. However, current biomedical MLLMs predominantly focus on image-level understanding and restrict interactions to textual commands, thus limiting their capability boundaries and the flexibility of usage. In this paper, we introduce a novel end-to-end multimodal large language model for the biomedical domain, named MedPLIB, which possesses pixel-level understanding. Excitingly, it supports visual question answering (VQA), arbitrary pixel-level prompts (points, bounding boxes, and free-form shapes), and pixel-level grounding. We propose a novel Mixture-of-Experts (MoE) multi-stage training strategy, which divides MoE into separate training phases for a visual-language expert model and a pixel-grounding expert model, followed by fine-tuning using MoE. This strategy effectively coordinates multitask learning while maintaining the computational cost at inference equivalent to that of a single expert model. To advance the research of biomedical MLLMs, we introduce the Medical Complex Vision Question Answering Dataset (MeCoVQA), which comprises an array of 8 modalities for complex medical imaging question answering and image region understanding. Experimental results indicate that MedPLIB has achieved state-of-the-art outcomes across multiple medical visual language tasks. More importantly, in zero-shot evaluations for the pixel grounding task, MedPLIB leads the best small and large models by margins of 19.7 and 15.6 respectively on the mDice metric.
DAMPER: A Dual-Stage Medical Report Generation Framework with Coarse-Grained MeSH Alignment and Fine-Grained Hypergraph Matching
PDF ↗Medical report generation is crucial for clinical diagnosis and patient management, summarizing diagnoses and recommendations based on medical imaging. However, existing work often overlook the clinical pipeline involved in report writing, where physicians typically conduct an initial quick review followed by a detailed examination. Moreover, current alignment methods may lead to misaligned relationships. To address these issues, we propose DAMPER, a dual-stage framework for medical report generation that mimics the clinical pipeline of report writing in two stages. In the first stage, a MeSH-Guided Coarse-Grained Alignment (MCG) stage that aligns chest X-ray (CXR) image features with medical subject headings (MeSH) features to generate a rough keyphrase representation of the overall impression. In the second stage, a Hypergraph-Enhanced Fine-Grained Alignment (HFG) stage that constructs hypergraphs for image patches and report annotations, modeling high-order relationships within each modality and performing hypergraph matching to capture semantic correlations between image regions and textual phrases. Finally,the coarse-grained visual features, generated MeSH representations, and visual hypergraph features are fed into a report decoder to produce the final medical report. Extensive experiments on public datasets demonstrate the effectiveness of DAMPER in generating comprehensive and accurate medical reports, outperforming state-of-the-art methods across various evaluation metrics.
Exposure correction aims to adjust the exposure of an under- and over-exposed image to enhance its overall visual quality. The core challenge of this task lies in that it requires to faithfully restore both the structure and perception information. In this work, we present a novel exposure correction method, referred to as CLIP-RestoreX, that leverages structural and perceptual priors from CLIP to tackle exposure correction. Specifically, we in CLIP-RestoreX propose to perform exposure correction by aligning CLIP-based structural and perceptual feature of the impaired image with its ground-truth image. To better restore the damaged structural information and perceptual information, we further design a frequency-domain based feature enhancement diffusion model, where we utilize the globality of Fourier transform to help reveal potential the relationship within the features. We conduct extensive experiments on several benchmark datasets. The results demonstrate that the proposed CLIP-RestoreX outperforms state-of-the-art exposure correction methods.
In few-shot action recognition (FSAR), long sub-sequences of video naturally express entire actions more effectively. However, the high computational complexity of mainstream Transformer-based methods limits their application. Recent Mamba demonstrates efficiency in modeling long sequences, but directly applying Mamba to FSAR overlooks the importance of local feature modeling and alignment. Moreover, long sub-sequences within the same class accumulate intra-class variance, which adversely impacts FSAR performance. To solve these challenges, we propose a Matryoshka MAmba and CoNtrasTive LeArning framework (Manta). Firstly, the Matryoshka Mamba introduces multiple Inner Modules to enhance local feature representation, rather than directly modeling global features. An Outer Module captures dependencies of timeline between these local features for implicit temporal alignment. Secondly, a hybrid contrastive learning paradigm, combining both supervised and unsupervised methods, is designed to mitigate the negative effects of intra-class variance accumulation. The Matryoshka Mamba and the hybrid contrastive learning paradigm operate in two parallel branches within Manta, enhancing Mamba for FSAR of long sub-sequence. Manta achieves new state-of-the-art performance on prominent benchmarks, including SSv2, Kinetics, UCF101, and HMDB51. Extensive empirical studies prove that Manta significantly improves FSAR of long sub-sequence from multiple perspectives.
Most indoor depth completion tasks rely on convolutional auto-encoders to reconstruct depth images, especially in areas with significant missing values. While traditional convolution treats valid and missing pixels equally, Partial Convolution (PConv) has mitigated this limitation. However, PConv fails to distinguish the varying degree of invalidity across different missing areas, which highlights the need for a more refined strategy. To solve this problem, we propose a novel system for indoor depth completion tasks that leverages Mask-adaptive Gated Convolution (MagaConv). MagaConv utilizes gated signals to selectively apply convolution kernels based on the characteristics of missing depth data. These gating signals are generated using shared convolution kernels that jointly process depth features and corresponding masks, ensuring coherent weight optimization. Additionally, the mask undergoes iterative updates according to predefined rules. To improve the fusion of depth and color information, we introduce a Bi-directional Aligning Projection (Bid-AP) module, which utilizes a bi-directional projection scheme with global spatial-channel attention mechanisms to filter out depth-irrelevant features from other modalities. Extensive experiments on popular benchmarks, including NYU-Depth V2, DIML, and SUN RGB-D, demonstrate that our model outperforms state-of-the-art methods in both accuracy and efficiency.
Dynamic 3D interaction has been attracting a lot of attention recently. However, creating such 4D content remains challenging. One solution is to animate 3D scenes with physics-based simulation, which requires manually assigning precise physical properties to the object or the simulated results would become unnatural. Another solution is to learn the deformation of 3D objects with the distillation of video generative models, which, however, tends to produce 3D videos with small and discontinuous motions due to the inappropriate extraction and application of physics priors. In this work, to combine the strengths and complementing shortcomings of the above two solutions, we propose to learn the physical properties of a material field with video diffusion priors, and then utilize a physics-based Material-Point-Method (MPM) simulator to generate 4D content with realistic motions. In particular, we propose motion distillation sampling to emphasize video motion information during distillation. In addition, to facilitate the optimization, we further propose a KAN-based material field with frame boosting. Experimental results demonstrate that our method enjoys more realistic motions than state-of-the-arts do.
Instance segmentation algorithms in remote sensing are typically based on conventional methods, limiting their application to seen scenarios and closed-set predictions. In this work, we propose a novel task called zero-shot remote sensing instance segmentation, aimed at identifying aerial objects that are absent from training data. Challenges arise when classifying aerial categories with high inter-class similarity and intra-class variance. Besides, the domain gap between vision-language models’ pretraining datasets and remote sensing datasets hinders the zero-shot capabilities of the pretrained model when it is directly applied to remote sensing images. To address these challenges, we propose a Zero-Shot Remote Sensing Instance Segmentation framework, dubbed ZoRI. Our approach features a discrimination-enhanced classifier that uses refined textual embeddings to increase the awareness of class disparities. Instead of direct fine-tuning, we propose a knowledge-maintained adaptation strategy that decouples semantic-related information to preserve vision-language alignment while adjusting features to capture remote sensing domain-specific visual cues. Additionally, we introduce a prior-injected prediction with cache bank of aerial visual prototypes to supplement the semantic richness of text embeddings and seamlessly integrate aerial representations, adapting to the remote sensing domain. We establish new experimental protocols and benchmarks, and extensive experiments demonstrate that ZoRI achieves the state-of-art performance on the zero-shot remote sensing instance segmentation task.
Unleashing the Temporal-Spatial Reasoning Capacity of GPT for Training-Free Audio and Language Referenced Video Object Segmentation
PDF ↗In this paper, we propose an Audio-Language-Referenced SAM 2 (AL-Ref-SAM 2) pipeline to explore the training-free paradigm for audio and language-referenced video object segmentation, namely AVS and RVOS tasks. The intuitive solution leverages GroundingDINO to identify the target object from a single frame and SAM 2 to segment the identified object throughout the video, which is less robust to spatiotemporal variations due to a lack of video context exploration. Thus, in our AL-Ref-SAM 2 pipeline, we propose a novel GPT-assisted Pivot Selection (GPT-PS) module to instruct GPT-4 to perform two-step temporal-spatial reasoning for sequentially selecting pivot frames and pivot boxes, thereby providing SAM 2 with a high-quality initial object prompt. Within GPT-PS, two task-specific Chain-of-Thought prompts are designed to unleash GPT’s temporal-spatial reasoning capacity by guiding GPT to make selections based on a comprehensive understanding of video and reference information. Furthermore, we propose a Language-Binded Reference Unification (LBRU) module to convert audio signals into language-formatted references, thereby unifying the formats of AVS and RVOS tasks in the same pipeline. Extensive experiments show that our training-free AL-Ref-SAM 2 pipeline achieves performances comparable to or even better than fully-supervised fine-tuning methods.
Personalized text-to-image generation methods can generate customized images based on the reference images, which have garnered wide research interest. Recent methods propose a finetuning-free approach with a decoupled cross-attention mechanism to generate personalized images requiring no test-time finetuning. However, when multiple reference images are provided, the current decoupled cross-attention mechanism encounters the object confusion problem and fails to map each reference image to its corresponding object, thereby seriously limiting its scope of application. To address the object confusion problem, in this work we investigate the relevance of different positions of the latent image features to the target object in diffusion model, and accordingly propose a weighted-merge method to merge multiple reference image features into the corresponding objects. Next, we integrate this weighted-merge method into existing pre-trained models and continue to train the model on a multi-object dataset constructed from the open-sourced SA-1B dataset. To mitigate object confusion and reduce training costs, we propose an object quality score to estimate the image quality for the selection of high-quality training samples. Furthermore, our weighted-merge training framework can be employed on single-object generation when a single object has multiple reference images. The experiments verify that our method achieves superior performance to the state-of-the-arts on the Concept101 dataset and DreamBooth dataset of multi-object personalized image generation, and remarkably improves the performance on single-object personalized image generation.
Face anti-spoofing (FAS) plays a pivotal role in ensuring the security and reliability of face recognition systems. With advancements in vision-language pretrained (VLP) models, recent two-class FAS techniques have leveraged the advantages of using VLP guidance, while this potential remains unexplored in one-class FAS methods. The one-class FAS focuses on learning intrinsic liveness features solely from live training images to differentiate between live and spoof faces. However, the lack of spoof training data can lead one-class FAS models to inadvertently incorporate domain information irrelevant to the live/spoof distinction (\eg, facial content), causing performance degradation when tested with a new application domain. To address this issue, we propose a novel framework called Spoof-aware one-class face anti-spoofing with Language Image Pretraining (SLIP). Given that live faces should ideally not be obscured by any spoof-attack-related objects (\eg, paper, or masks) and are assumed to yield zero spoof cue maps, we first propose an effective language-guided spoof cue map estimation to enhance one-class FAS models by simulating whether the underlying faces are covered by attack-related objects and generating corresponding nonzero spoof cue maps. Next, we introduce a novel prompt-driven liveness feature disentanglement to alleviate live/spoof-irrelative domain variations by disentangling live/spoof-relevant and domain-dependent information. Finally, we design an effective augmentation strategy by fusing latent features from live images and spoof prompts to generate spoof-like image features and thus diversify latent spoof features to facilitate the learning of one-class FAS. Our extensive experiments and ablation studies support that SLIP consistently outperforms previous one-class FAS methods.
VProChart: Answering Chart Question Through Visual Perception Alignment Agent and Programmatic Solution Reasoning
PDF ↗Charts are widely used for data visualization across various fields, including education, research, and business. Chart Question Answering (CQA) is an emerging task focused on the automatic interpretation and reasoning of data presented in charts. However, chart images are inherently difficult to interpret, and chart-related questions often involve complex logical and numerical reasoning, which hinders the performance of existing models. This paper introduces VProChart, a novel framework designed to address these challenges in CQA by integrating a lightweight Visual Perception Alignment Agent (VPAgent) and a Programmatic Solution Reasoning approach. VPAgent aligns and models chart elements based on principles of human visual perception, enhancing the understanding of chart context. The Programmatic Solution Reasoning approach leverages large language models (LLMs) to transform natural language reasoning questions into structured solution programs, facilitating precise numerical and logical reasoning. Extensive experiments on benchmark datasets such as ChartQA and PlotQA demonstrate that VProChart significantly outperforms existing methods, highlighting its capability in understanding and reasoning with charts.
Chart understanding enables automated data analysis for humans, which requires models to achieve highly accurate visual comprehension. While existing Visual Language Models (VLMs) have shown progress in chart understanding, the lack of high-quality training data and comprehensive evaluation benchmarks hinders VLM chart comprehension. In this paper, we introduce EvoChart, a novel self-training method for generating synthetic chart data to enhance VLMs' capabilities in real-world chart comprehension. We also propose EvoChart-QA, a noval benchmark for measuring models' chart comprehension abilities in real-world scenarios. Specifically, EvoChart is a unique self-training data synthesis approach that simultaneously produces high-quality training corpus and a high-performance chart understanding model. EvoChart-QA consists of 650 distinct real-world charts collected from 140 different websites and 1,250 expert-curated questions that focus on chart understanding. Experimental results on various open-source and proprietary VLMs tested on EvoChart-QA demonstrate that even the best proprietary model, GPT-4o, achieves only 49.8% accuracy. Moreover, the EvoChart method significantly boosts the performance of open-source VLMs on real-world chart understanding tasks, achieving 54.2% accuracy on EvoChart-QA.
AUTE: Peer-Alignment and Self-Unlearning Boost Adversarial Robustness for Training Ensemble Models
PDF ↗Adversarial attacks poses a significant threat to the security of AI-based systems. To counteract these attacks, adversarial training (AT) and ensemble learning (EL) have emerged as widely adopted methods for enhancing model robustness. However, a counter-intuitive phenomenon arises where the simple combination of these approaches may potentially compromising adversarial robustness of ensemble models. In this paper, we propose a novel method called Alignment and Unlearning for Training Ensembles (AUTE), aiming to effectively integrate AT and EL to maximize their benefits. Specifically, AUTE incorporates two key components. Firstly, AUTE divides the ensemble into a big peer model and a single member in a loop manner, aligning their outputs for boosting robustness of each member. Secondly, AUTE introduces the concept of unlearning, actively forgetting specific data with over-confident properties to preserve model capacity to learn more robust features. Extensive experiments across various datasets and networks illustrate that AUTE achieves superior performance compared to baselines. For instance, a 5-member AUTE with ResNet-20 networks outperforms state-of-the-art method by 2.1% and 3.2% in classifying clean and adversarial data. Additionally, AUTE can easily extend to non-adversarial training paradigm, surpassing current standard ensemble learning methods by a large margin.
Pansharpening aims to combine a high-resolution panchromatic (PAN) image with a low-resolution multispectral (LRMS) image to produce a high-resolution multispectral (HRMS) image. Although pansharpening in the frequency domain offers clear advantages, most existing methods either continue to operate solely in the spatial domain or fail to fully exploit the benefits of the frequency domain. To address this issue, we innovatively propose Multi-Frequency Fusion Attention (MFFA), which leverages wavelet transforms to cleanly separate frequencies and enable lossless reconstruction across different frequency domains. Then, we generate Frequency-Query, Spatial-Key, and Fusion-Value based on the physical meanings represented by different features, which enables a more effective capture of specific information in the frequency domain. Additionally, we focus on the preservation of frequency features across different operations. On a broader level, our network employs a wavelet pyramid to progressively fuse information across multiple scales. Compared to previous frequency domain approaches, our network better prevents confusion and loss of different frequency features during the fusion process. Quantitative and qualitative experiments on multiple datasets demonstrate that our method outperforms existing approaches and shows significant generalization capabilities for real-world scenarios.
In the domain of computer vision, Parameter-Efficient Tuning (PET) is increasingly replacing the traditional paradigm of pre-training followed by full fine-tuning. PET is particularly favored for its effectiveness in large foundation models, as it streamlines transfer learning costs and optimizes hardware utilization. However, the current PET methods are mainly designed for single-modal optimization. While some pioneering studies have undertaken preliminary explorations, they still remain at the level of aligned encoders (e.g., CLIP) and lack exploration of misaligned encoders. These methods show sub-optimal performance with misaligned encoders, as they fail to effectively align the multimodal features during fine-tuning. In this paper, we introduce DETRIS, a parameter-efficient tuning framework designed to enhance low-rank visual feature propagation by establishing dense interconnections between each layer and all preceding layers, which enables effective cross-modal feature interaction and adaptation to misaligned encoders. We also suggest using text adapters to improve textual features. Our simple yet efficient approach greatly surpasses state-of-the-art methods with 0.9% to 1.8% backbone parameter updates, evaluated on challenging benchmarks.