CVPR 2026 论文列表

Recent advances in video large language models have demonstrated strong capabilities in understanding short clips. However, scaling them to hours- or days-long videos remains highly challenging due to limited context capacity and the loss of critical visual details during abstraction. Existing memory-augmented methods mitigate this by leveraging textual summaries of video segments, yet they heavily rely on text and fail to utilize visual evidence when reasoning over complex scenes. Moreover, retrieving from fixed temporal scales further limits their flexibility in capturing events that span variable durations. To address this, we introduce WorldMM, a novel multimodal memory agent that constructs and retrieves from multiple complementary memories, encompassing both textual and visual representations. WorldMM comprises three types of memory: episodic memory indexes factual events across multiple temporal scales, semantic memory continuously updates high-level conceptual knowledge, and visual memory preserves detailed information about scenes. During inference, an adaptive retrieval agent iteratively selects the most relevant memory source and leverages multiple temporal granularities based on the query, continuing until it determines that sufficient information has been gathered. WorldMM significantly outperforms existing baselines across five long video question-answering benchmarks, achieving an average 8.4% performance gain over previous state-of-the-art methods, showing its effectiveness on long video reasoning.

Backdoor attacks pose a significant threat to deep neural networks. Recent studies have shown that model compression, such as quantization and pruning, can be exploited by attackers to implant conditional backdoors. Such backdoors remain dormant in the original model but are activated after the model undergoes specific operations, making them highly stealthy and difficult to detect. Traditional defense methods struggle to counter this type of attack, while defenses specifically designed for conditional backdoors also have difficulty handling traditional backdoor attacks. To address these challenges, we propose a universal defense method, termed Logit Margin Repulsion (LMR). LMR uses a small set of clean samples and combines selective cross-entropy with a logit-margin constraint to enlarge the gap between the backdoor class and benign classes. It then removes channels associated with backdoor behavior through selective pruning, thereby achieving strong backdoor purification. Extensive experiments on a variety of CNNs and Vision Transformers demonstrate that, even with an extremely limited amount of clean data (0.1%), LMR can effectively mitigate both traditional and conditional backdoor attacks. The implementation is publicly available on https://github.com/Trusted-LLM/LMR.

Medical imaging remains challenging due to acoustic shadows, motion blur, and indistinct boundaries, while adapting vision models to such domains often requires heavy task-specific fine-tuning and can degrade general-image capability. We propose DCRM-ViT, a domain-conditioned residual modulation framework for Vision Transformers that preserves general-vision knowledge while adapting to diverse medical and natural domains. DCRM-ViT keeps the backbone frozen and augments each block with lightweight Residual Modulation Blocks (RMBs), whose parameters are synthesized per sample by a Domain Router (DR) and a Parameter Synthesizer Network (PSN). The DR predicts soft domain weights from input features, and the PSN maps them to low-rank residuals that modulate selected projections and optionally add a domain-aware attention bias. We train the model with a bi-level optimization scheme, where an inner loop adapts RMBs to task supervision and an outer loop updates DR, PSN, and RMB initialization to improve cross-domain generalization. Across fine-grained classification (Food101, SUN397, Stanford Cars) and medical segmentation (ultrasound, CT, MRI), DCRM-ViT consistently outperforms strong baselines with modest trainable compute. Ablation studies confirm the contribution of the proposed components. Overall, DCRM-ViT achieves strong cross-domain performance with low overhead, requiring only 4.7 GFLOPs and 0.3 min/epoch.

Missing modalities in RGBT tracking often lead to incomplete and unstable multimodal feature representations that greatly degrade the performance. Existing methods typically attempt to recover missing modalities from available ones, but the quality of data generated in challenging scenarios might be unsatisfactory. In addition, current approaches exhibit limited flexibility in processing both missing and complete data. To overcome these limitations, we propose a Spatio-temporal Conditional Denoising Transformer (SCDT), which integrates the spatial cues and the temporal context to adaptively perform information reconstruction of missing modalities and feature enhancement of weak modalities in a unified framework, for robust modality-missing RGBT tracking. In particular, SCDT leverages the short-term temporal cues from recent historical frames to capture the fine-grained temporal correlations and the long-term temporal cues encoding modality evolution to capture the global context. By jointly exploiting long short-term temporal contexts as the conditions, SCDT progressively guides noisy features of available modalities to learn reliable and temporally consistent multimodal representations. Furthermore, SCDT introduces a noise-modulated adaptation mechanism that dynamically adjusts its behavior according to the modal availability, enabling a single framework to unify feature learning under both modality-missing and complete scenarios without changing the architecture or parameters. Extensive experiments on three public benchmark datasets demonstrate that our method consistently outperforms state-of-the-art methods.

Accurate estimation of food nutrition plays a vital role in promoting healthy dietary habits and personalized diet management. Most existing food datasets primarily focus on Western cuisines and lack sufficient coverage of Chinese dishes, which restricts accurate nutritional estimation for Chinese meals. Moreover, many state-of-the-art nutrition prediction methods rely on depth sensors, restricting their applicability in daily scenarios. To address these limitations, we introduce OmniFood8K, a comprehensive multimodal dataset comprising 8,036 food samples, each with detailed nutritional annotations and multi-view images. In addition, to enhance models' capability in nutritional prediction, we construct NutritionSynth-115K, a large-scale synthetic dataset that introduces compositional variations while preserving precise nutritional labels. Moreover, we propose an end-to-end framework for nutritional prediction from a single RGB image. First, we predict a depth map from a single RGB image and design the Scale-Shift Residual Adapter (SSRA) to refine it for global scale consistency and local structural preservation. Second, we propose the Frequency-Aligned Fusion Module (FAFM) to hierarchically align and fuse RGB and depth features in the frequency domain. Finally, we design a Mask-based Prediction Head (MPH) to emphasize key ingredient regions via dynamic channel selection for more accurate prediction. Extensive experiments on multiple datasets demonstrate the superiority of our method over existing approaches. Project homepage: https://yudongjian.github.io/OmniFood8K-food/

Dynamic data pruning accelerates deep learning by selectively omitting less informative samples during training. While per-sample loss is a common importance metric, obtaining it can be challenging or infeasible for complex models or loss functions, often requiring significant implementation effort. This work proposes the Batch Loss Score (BLS), a computationally efficient alternative using an Exponential Moving Average (EMA) of readily available batch losses to assign scores to individual samples. We frame the batch loss, from the perspective of a single sample, as a noisy measurement of its scaled individual loss, with noise originating from stochastic batch composition. It is formally shown that the EMA mechanism functions as a first-order low-pass filter, attenuating high-frequency batch composition noise. This yields a score approximating the smoothed and persistent contribution of the individual sample to the loss, providing a theoretical grounding for BLS as a proxy for sample importance. BLS demonstrates remarkable code integration simplicity (three-line injection) and readily adapts existing per-sample loss-based methods (one-line proxy). Its effectiveness is demonstrated by enhancing two such methods to losslessly prune 20%-50% of samples across 14 datasets, 11 tasks and 18 models, highlighting its utility and broad applicability, especially for complex scenarios where per-sample loss is difficult to access. Code is available at https://github.com/mrazhou/BLS.

Timely and accurate forecasts of severe weather events are essential for early warning and for constraining downstream analysis and decision-making. Since severe weather events prediction still depends on subjective, time-consuming expert interpretation, end-to-end "AI weather station" systems are emerging but face three major challenges: (1) scarcity of severe weather event samples; (2) imperfect alignment between high-dimensional meteorological data and textual warnings; (3) current multimodal language models cannot effectively process high-dimensional meteorological inputs or capture their complex spatiotemporal dependencies. To address these challenges, we introduce MP-Bench, the first large-scale multimodal dataset for severe weather events prediction, comprising 421,363 pairs of raw multi-year meteorological data and corresponding text caption, covering a wide range of severe weather scenarios. On top of this dataset, we develop a Meteorology Multimodal Large Model (MMLM) that directly ingests 4D meteorological inputs. In addition, it is designed to accommodate the unique characteristics of 4D meteorological data flow, incorporating three plug-and-play adaptive fusion modules that enable dynamic feature extraction and integration across temporal sequences, vertical pressure layers, and spatial dimensions. Extensive experiments on MP-Bench show that MMLM achieves strong performance across multiple tasks, demonstrating effective severe weather understanding and representing a key step toward automated, AI-driven severe weather events forecasting systems.

Vision-Language Models like CLIP are extensively used for inter-modal tasks which involve both visual and text modalities. However, when the individual modality encoders are applied to inherently intra-modal tasks like image-to-image retrieval, their performance suffers from the intra-modal misalignment. In this paper we study intra-modal misalignment in CLIP with a focus on the role of the projectors that map pre-projection image and text embeddings into the shared embedding space. By analyzing the form of the cosine similarity applied to projected features, and its interaction with the contrastive CLIP loss, we show that there is an inter-modal operator responsible for aligning the two modalities during training, and a second, intra-modal operator that only enforces intra-modal normalization but does nothing to promote intra-modal alignment. Via spectral analysis of the inter-modal operator, we identify an approximately isotropic subspace in which the two modalities are well-aligned, as well as anisotropic directions specific to each modality. We demonstrate that this aligned subspace can be directly obtained from the projector weights and that removing the anisotropic directions improves intra-modal alignment. Our experiments on intra-modal retrieval and classification benchmarks show that our training-free method reduces intra-modal misalignment, greatly lowers latency, and outperforms existing approaches across multiple pre-trained CLIP-like models. The code is publicly available at: https://github.com/simomagi/IsoCLIP.

Structured shape completion recovers missing geometry as primitives rather than as unstructured points, which enables primitive-based surface reconstruction. Instead of following the prevailing cascade, we rethink how primitives and points should interact, and find it more effective to decode primitives in a dedicated pathway that attends to shared shape features. Following this principle, we present UniCo, which in a single feed-forward pass predicts a set of primitives with complete geometry, semantics, and inlier membership. To drive this unified representation, we introduce primitive proxies, learnable queries that are contextualized to produce assembly-ready outputs. To ensure consistent optimization, our training strategy couples primitives and points with online target updates. Across synthetic and real-world benchmarks with four independent assembly solvers, UniCo consistently outperforms recent baselines, lowering Chamfer distance by up to 50% and improving normal consistency by up to 7%. These results establish an attractive recipe for structured 3D understanding from incomplete data. Project page: https://unico-completion.github.io.

Multimodal large language models (MLLMs) have demonstrated impressive reasoning and instruction-following capabilities, yet their expanded modality space introduces new compositional safety risks that emerge from complex text-image interactions.Such cross-modal couplings can produce unsafe semantics even when individual inputs are benign, exposing the fragile safety awareness of current MLLMs.While recent works enhance safety by guiding models to reason about potential risks, unregulated reasoning traces may compromise alignment; although Group Relative Policy Optimization (GRPO) offers self-rewarded refinement without human supervision, it lacks verifiable signals for reasoning safety.To address this, we propose **SafeGRPO** a self-rewarded multimodal safety alignment framework that integrates rule-governed reward construction into GRPO, enabling interpretable and verifiable optimization of reasoning safety. Built upon the constructed **SafeTag-VL-3K** dataset with explicit visual, textual, and combined safety tags, SafeGRPO performs **step-guided safety thinking** to enforce structured reasoning and behavior alignment, substantially improving multimodal safety awareness, compositional robustness, and reasoning stability across diverse benchmarks without sacrificing general capabilities.

Universal Multimodal Embeddings (UMEs) aim to unify various modalities and tasks into a shared representation space. In recent years, this field has witnessed substantial progress driven by the development of Multimodal Large Language Models (MLLMs). However, a crucial capability, visual identity discrimination, remains underexplored in existing UME methods, despite its critical role in a wide range of tasks, including instance retrieval, re-identification, and identity preservation in AI-generated content. To bridge this gap, we propose a unified formulation for visual identity discrimination (VisID) and introduce MVEB (Multimodal Visual Identity Embedding Benchmark), a large-scale benchmark curated from both real-world and synthetic datasets to support evaluation and training. Furthermore, we present a simple yet effective learning framework that jointly optimizes general multimodal and visual identity representations through a carefully designed identity-aware sampling mechanism. Extensive experiments demonstrate that our approach successfully endows UMEs with strong identity discrimination capability and maintains competitive general multimodal performance. We believe this work not only illuminates a critical yet neglected capability, but also takes a step toward more holistic universal multimodal embeddings. Code and data are available at \href https://chrisclear3.github.io/MVEB MVEB .

While powerful in image-conditioned generation, multimodal large language models (MLLMs) can display uneven performance across demographic groups, highlighting fairness risks. In safety-critical clinical settings, such disparities risk producing unequal diagnostic narratives and eroding trust in AI-assisted decision-making. While fairness has been studied extensively in vision-only and language-only models, its impact on MLLMs remains largely underexplored. To address these biases, we introduce FairLLaVA, a parameter-efficient fine-tuning method that mitigates group disparities in visual instruction tuning without compromising overall performance. By minimizing the mutual information between target attributes, FairLLaVA regularizes the model's representations to be demographic-invariant. The method can be incorporated as a lightweight plug-in, maintaining efficiency with low-rank adapter fine-tuning, and provides an architecture-agnostic approach to fair visual instruction following. Extensive experiments on large-scale chest radiology report generation and dermoscopy visual question answering benchmarks show that FairLLaVA consistently reduces inter-group disparities while improving both equity-scaled clinical performance and natural language generation quality across diverse medical imaging modalities. Code can be accessed at https://github.com/bhosalems/FairLLaVA.

Traditional open-access datasets focusing on surgical procedures are often limited by their small size, typically consisting of fewer than 100 videos and less than 30 hours of footage, which leads to poor model generalization. To address this data limitation, a new dataset called LEMON has been compiled using a novel aggregation pipeline that collects high-resolution videos from online sources. Featuring an extensive collection of over 4K surgical videos totaling 938 hours (85 million frames) of high-quality footage across multiple procedure types, LEMON offers a comprehensive resource surpassing existing alternatives in size and scope, including two novel downstream tasks. To demonstrate the effectiveness of this diverse dataset, we introduce LemonFM, a foundation model pretrained on LEMON using a novel self-supervised augmented knowledge distillation approach. LemonFM consistently outperforms existing surgical foundation models across four downstream tasks and six datasets, achieving significant gains in surgical phase recognition (+9.5pp, +9.4pp, and +8.4pp in Jaccard on AutoLaparo, M2CAI16, and Cholec80), surgical action recognition (+4.4pp in mAP on CholecT50), surgical tool presence detection (+5.3pp and +10.2pp in mAP on Cholec80 and GraSP), and surgical semantic segmentation (+10.3pp in mDice on CholecSeg8k). LEMON and LemonFM will serve as foundational resources for the research community and industry, accelerating progress in developing autonomous robotic surgery systems and ultimately contributing to safer and more accessible surgical care worldwide. Dataset, code, and models are publicly available at https://github.com/visurg-ai/LEMON.

Crop yield prediction requires substantial data to train scalable models. However, creating yield prediction datasets is constrained by high acquisition costs, heterogeneous data quality, and data privacy regulations. Consequently, existing datasets are scarce, low in quality, or limited to regional levels or single crop types, hindering the development of scalable data-driven solutions. In this work, we release YieldSAT, a large, high-quality, and multimodal dataset for high-resolution crop yield prediction. YieldSAT spans various climate zones across multiple countries, including Argentina, Brazil, Uruguay, and Germany, and includes major crop types, including corn, rapeseed, soybeans, and wheat, across 2,173 expert-curated fields. In total, over 12.2 million yield samples are available, each with a spatial resolution of 10 m. Each field is paired with multispectral satellite imagery, resulting in 113,555 labeled satellite images, complemented by auxiliary environmental data. We demonstrate the potential of large-scale and high-resolution crop yield prediction as a pixel regression task by comparing various deep learning models and data fusion architectures. Furthermore, we highlight open challenges arising from severe distribution shifts in the ground truth data under real-world conditions. To mitigate this, we explore a domain-informed Deep Ensemble approach that exhibits significant performance gains. The dataset is available at https://yieldsat.github.io/.

Multimodal large language models (MLLMs) have achieved remarkable performance across a wide range of vision-language tasks. However, their ability in low-level visual perception, particularly in detecting fine-grained visual discrepancies, remains underexplored and lacks systematic analysis.In this work, we introduce OddGridBench, a controllable benchmark for evaluating the visual discrepancy sensitivity of MLLMs. OddGridBench comprises over 1,400 grid-based images, where a single element differs from all others by one or multiple visual attributes such as color, size, rotation, or position.Experiments reveal that all evaluated MLLMs, including open-source families such as Qwen3-VL and InternVL3.5, and proprietary systems like Gemini-2.5-Pro and GPT-5, perform far below human levels in visual discrepancy detection. We further propose OddGrid-GRPO, a reinforcement learning framework that integrates curriculum learning and distance-aware reward. By progressively controlling the difficulty of training samples and incorporating spatial proximity constraints into the reward design, OddGrid-GRPO significantly enhances the model's fine-grained visual discrimination ability.We hope OddGridBench and OddGrid-GRPO will lay the groundwork for advancing perceptual grounding and visual discrepancy sensitivity in multimodal intelligence.All resources will be publicly released upon acceptance.

Image denoising is a fundamental task in computer vision aimed at recovering clean images from noise-corrupted observations. While supervised deep learning methods achieve remarkable performance when trained on paired data with known noise levels, their real-world applicability is limited as noise characteristics are often unknown. Existing unsupervised techniques, such as blind-spot networks or methods based on statistical estimation, either compromise performance due to information loss or suffer from inaccuracies in noise level estimation. To address these challenges, we propose a novel two-stage self-supervised denoising framework that first accurately estimates the noise level directly from noisy images, without requiring clean references or prior noise knowledge. Building upon theoretical insights from Noisier2Noise, we rigorously derive a relationship between the noise level and the variance of the denoised image, enabling robust estimation via a deep learning model and a ternary search strategy. The estimated noise level is then used to synthesize training pairs for supervised denoising. Experiments demonstrate that our method outperforms existing unsupervised approaches and traditional noise estimation techniques, achieving performance competitive with--and in some cases surpassing--supervised methods trained with known noise levels. The proposed framework effectively overcomes the training data pair limitations of supervised approaches for unknown additive white Gaussian noise. Our code is available at https://github.com/zhanzhanblue/CANC.

Multimodal Sentiment Analysis (MSA) aims to integrate textual, acoustic, and visual information to predict sentiment polarity. With the emergence of Large Language Models (LLMs), existing studies commonly employ learnable queries to compress audio-visual representations and feed them as soft prompts into LLMs for MSA. However, due to the implicit learning mechanism of the learnable queries, these learnable queries lack explicit guidance regarding how each query encodes sentiment semantics. To address this issue, we propose a prototype-as-prompt framework that maps audio-visual representations into a fixed set of multimodal sentiment prototypes. These prototypes are then used as soft prompts to guide the LLM in performing MSA. Concretely, we first compress both textual and non-textual features into multimodal prototypes using a resampling-based strategy. We further introduce a sentiment-aware prototype learning that explicitly binds multimodal prototypes with sentiment semantics. To ensure both cross-modal consistency and intra-modal diversity of multimodal sentiment prototypes, we design a cross-modal prototype alignment constraint and a distance-weighted prototype diversity constraint. Extensive experiments across three LLMs and four benchmark datasets show that PaP achieves superior performance with only 0.09%-0.26% of trainable parameters, highlighting its effectiveness and parameter efficiency.

Modern mapping tools remain fundamentally point-and-click systems, offering limited support for interactive, multimodal engagement with either map views or real-world camera inputs. This limitation becomes most apparent in the last 100 meters of navigation and during in-situ local discovery, where users naturally combine spatial reasoning over maps with egocentric perception of their surroundings. We introduce IMAIA, an interactive geospatial assistance framework that unifies three core capabilities: (1) map-centric spatial understanding, (2) camera-to-place grounding, and (3) human-centered, bearing-aware navigation. IMAIA consists of two interoperable components--Maps Plus and PAISA--coordinated by a lightweight multi-agent orchestrator. Maps Plus converts vector and satellite maps into a grid-aligned representation that enables flexible, view-conditioned spatial queries. PAISA fuses camera imagery with geospatial signals such as location, heading, and proximity to interpret egocentric scenes and surface relevant attributes of nearby places. The framework is fully modular: vision-language backends can be replaced without altering system behavior. IMAIA enables responsive, interpretable, and context-aware assistance across both map-based and real-world interactions, supporting a new class of interactive mapping experiences that are practical for real-world deployment.

Occluded visible-infrared person re-identification (Occluded VI-ReID) remains difficult due to modality heterogeneity and occlusions, both of which break structural consistency and weaken cross-modality feature alignment. Existing methods rely mainly on spatial-domain cues (such as local body parts and salient patches), but their discriminability degrades severely under varying imaging conditions or partial visibility. To address these issues, we introduce a spatial-frequency collaborative perspective that offers global perception and cross-location consistency. Specifically, we propose a Spatial-Frequency Collaborative Learning (SFCL) framework that uses frequency information to complement spatial representations. SFCL comprises a Cross-Modality Frequency Alignment Module (CFAM), a Spatial-Frequency Interaction Module (SFIM), and a Frequency-Aware Discriminative (FAD) loss. The CFAM models the spectral features of visible/infrared images in the frequency domain, establishing modality-consistent spectral priors. The SFIM injects these priors into spatial features, promoting dual-domain interaction and complementary representations of spatial and frequency semantics. In addition, the FAD loss jointly enforces cross-modality frequency alignment and semantic consistency, thus enhancing robustness and discriminability under occlusions. For real-occlusion evaluation, we construct two occluded datasets, Occ-SYSU-MM01 and Occ-RegDB, on which SFCL outperforms the state-of-the-art.