CVPR 2026 论文列表

Conventional cameras generate a lot of data that can be challenging to process in resource-constrained applications. Usually, cameras generate data streams on the order of the number of pixels in the image. However, most of this captured data is redundant for many downstream computer vision algorithms. We propose a novel camera design, which we call SuperCam, that adaptively processes captured data by performing superpixel segmentation on the fly. We show that SuperCam performs better than current state-of-the-art superpixel algorithms under memory-constrained situations. We also compare how well SuperCam performs when the compressed data is used for downstream computer vision tasks. Our results demonstrate that the proposed design provides superior output for image segmentation, object detection, and monocular depth estimation in situations where the available memory on the camera is limited. We posit that superpixel segmentation will play a crucial role as more computer vision inference models are deployed in edge devices. SuperCam would allow computer vision engineers to design more efficient systems for these applications.

We present BioCoach, a biomechanics-grounded vision-language framework for fitness coaching from streaming video. BioCoach fuses two signals, visual appearance and 3D skeletal kinematics, through a novel three-stage pipeline: an exercise-specific degree-of-freedom selector that focuses analysis on salient joints; a structured biomechanical context that pairs individualized morphometrics with cycle and constraint analysis; and a vision-biomechanics conditioned feedback module that applies cross-attention to generate precise, actionable text. Using parameter-efficient training that freezes the vision and language backbones, BioCoach yields transparent, personalized reasoning rather than pattern matching. To enable learning and fair evaluation, we augment QEVD-fit-coach with biomechanics-oriented feedback to create QEVD-bio-fit-coach, and we introduce a biomechanics-aware LLM judge metric. BioCoach delivers clear gains on QEVD-bio-fit-coach across lexical and judgment metrics while maintaining temporal triggering; on the original QEVD-fit-coach, it improves text quality and correctness with near-parity timing, demonstrating that explicit kinematics and constraints are key to accurate, phase-aware coaching.

While model merging has demonstrated remarkable success for large language models (LLMs), its application to vision-language models (VLMs) remains largely underexplored. Recent methods attempt to enhance VLM reasoning capabilities by integrating specialized LLM parameters through layer-wise merging. However, existing paradigms suffer from two critical limitations: (1) strict positional correspondence, enforcing rigid one-to-one layer alignment, and (2) uniform merging weights applied across all layers. These constraints fail to account for substantial functional disparities between VLM and LLM layers, potentially misaligning incompatible layers and leading to detrimental parameter combinations. To address these, we propose Chain-of-Merging (CoM), a framework that adaptively adjusts merging plans for different images and questions, comprising two key stages: (1) Adaptive Layer Matching, which identifies optimal layer pairings based on structural and semantic matching scores while filtering incompatible pairings, and (2) Dynamic Weight Merging, which determines layer-specific merging weights based on matching scores and employs spherical linear interpolation to minimize memory overhead. Extensive experiments demonstrate that CoM achieves substantial performance improvements, with Qwen2.5-VL-7B + Qwen2.5-Math-7B attaining a 4.4% average improvement on mathematical reasoning benchmarks while enhancing general visual understanding, significantly outperforming existing training-free methods.

Semantic Scene Completion (SSC) is essential for 3D perception in mobile robotics, as it enables holistic scene understanding by jointly estimating dense volumetric occupancy and per-voxel semantics. Although SSC has been widely studied in terrestrial domains such as autonomous driving, aerial settings like autonomous flying remain largely unexplored, thereby limiting progress on downstream applications. Furthermore, LiDAR sensors are the primary modality for SSC data generation, which poses challenges for most uncrewed aerial vehicles (UAVs) due to flight regulations, mass and energy constraints, and the sparsity of LiDAR point clouds from elevated viewpoints. To address these limitations, we propose a LiDAR-free, camera-based data generation framework. By leveraging classical 3D reconstruction, our framework automates semantic label transfer by lifting <10% of annotated images into the reconstructed point cloud, substantially minimizing manual 3D annotation effort. Based on this framework, we introduce OccuFly, the first real-world, camera-based aerial SSC benchmark, captured across multiple altitudes and all seasons. OccuFly provides over 20,000 samples of images, semantic voxel grids, and metric depth maps across 21 semantic classes in urban, industrial, and rural environments, and follows established data organization for seamless integration. We benchmark both SSC and metric monocular depth estimation on OccuFly, revealing fundamental limitations of current vision foundation models in aerial settings and establishing new challenges for robust 3D scene understanding in the aerial domain. Visit https://github.com/markus-42/occufly.

We present AToken, the first unified visual tokenizer that achieves both high-fidelity reconstruction and semantic understanding across images, videos, and 3D assets. Unlike existing tokenizers that specialize in either reconstruction or understanding for single modalities, AToken encodes these diverse visual inputs into a shared 4D latent space, unifying both tasks and modalities in a single framework. Specifically, we introduce a pure transformer architecture with 4D rotary position embeddings to process visual inputs of arbitrary resolutions and temporal durations. To ensure stable training, we introduce an adversarial-free training objective that combines perceptual and Gram matrix losses, achieving state-of-the-art reconstruction quality. By employing a progressive training curriculum, AToken gradually expands from single images, videos, and 3D, and supports both continuous and discrete latent tokens. AToken achieves 0.21 rFID with 82.2% ImageNet accuracy for images, 3.01 rFVD with 40.2% MSRVTT retrieval for videos, and 28.28 PSNR with 90.9% classification accuracy for 3D.. In downstream applications, AToken enables both visual generation tasks (e.g., image generation with continuous and discrete tokens, text-to-video generation, image-to-3D synthesis) and understanding tasks (e.g., multimodal LLMs), achieving competitive performance across all benchmarks. These results shed light on the next-generation multimodal AI systems built upon unified visual tokenization.

The proliferation of low-power intelligent processors with integrated Neural Processing Units (NPUs), called mNPUs, has created new opportunities for on-device generative AI, benefitting end devices like smart wearables and small robots. However, deploying Vision-Language Models (VLMs) on mNPUs is severely hindered by stringent memory constraints and limited operator support. To bridge this critical gap, we propose mVLM, the first lightweight-oriented VLM architecture designed for mNPUs. It is comprised of our proposed OverMod encoder and AttSSM decoder. OverMod is a lightweight dynamic convolutional network inspired by biomimetic vision, incorporating our novel Global Spatial Modulation mechanism to enable adaptive, high-fidelity feature extraction using only NPU-friendly operators. AttSSM leverages a highly efficient State Space Model (SSM) core, augmented with multi-scale feature fusion and Global Context Dynamic Modulation mechanism, to perform robust sequential modeling. Furthermore, we introduce a coordinated full-parameter quantization strategy that preserves precision across the encoder-decoder boundary, alongside hand-optimized operators for unsupported modules like SSMs. mVLM achieves a competitive CIDEr score of 117.8 on the COCO Karpathy test split and, for the first time, demonstrates the feasibility of millisecond-level VLM inference on a mNPU platform.

While Instruction-based Image Editing (IIE) has achieved significant progress, existing benchmarks pursue task breadth via mixed evaluations. This paradigm obscures a critical failure mode crucial in professional applications: the inconsistent performance of models across tasks of varying semantic scales. To address this gap, we introduce Omni IIE Bench, a high-quality, human-annotated benchmark specifically designed to diagnose the editing consistency of IIE models in practical application scenarios. Omni IIE Bench features an innovative dual-track diagnostic design: (1) Single-turn Consistency, comprising shared-context task pairs of attribute modification and entity replacement; and (2) Multi-turn Coordination, involving continuous dialogue tasks that traverse semantic scales. The benchmark is constructed via an exceptionally rigorous multi-stage human filtering process, incorporating a quality standard enforced by computer vision graduate students and an industry relevance review conducted by professional designers. We perform a comprehensive evaluation of 8 mainstream IIE models using Omni IIE Bench. Our analysis quantifies, for the first time, a prevalent performance gap: nearly all models exhibit a significant performance degradation when transitioning from low-semantic-scale to high-semantic-scale tasks. Omni IIE Bench provides critical diagnostic tools and insights for the development of next-generation, more reliable, and stable IIE models.

Large Vision-Language Models (VLMs) enable strong multimodal reasoning but incur heavy inference costs from redundant visual tokens. Token pruning alleviates this issue, yet existing approaches face limitations. Attention-based methods rely on raw attention scores, which are often unstable across layers and heads and can lead to redundant selections. Diversity-based methods improve robustness by selecting tokens far apart in feature space, but risk dropping regions needed for accurate prediction. We propose ZOO-Prune, a training-free framework built on the intuition that highly sensitive tokens have a stronger influence on the model's output and capture complementary visual cues rather than redundant ones. To achieve this, we estimate token sensitivity using zeroth-order perturbations at the lightweight projection layer. This measures how small random perturbations affect the projected features and enables efficient approximation of each token's influence without backpropagation. Extensive experiments across multiple VLMs and benchmarks show that ZOO-Prune consistently outperforms prior methods while pruning up to 94.4% of tokens without sacrificing accuracy. Our method also improves efficiency, reaching up to 2.30x faster end-to-end inference compared to the baseline. Code is available at https://aim-skku.github.io/ZOO-Prune.

In the deep-learning-based computer vision community, Neural Architecture Search (NAS) has become the de-facto tool for acquiring task-optimal network structures. Nevertheless, NAS methods are trapped in a fundamental accuracy-efficiency dilemma: training-based approaches deliver reliable performance but incur prohibitive search costs, whereas training-free strategies are ultra-fast but often yield relatively unreliable rankings. To reconcile this conflict, we propose a vision-oriented lightweight training-based NAS framework. We first design six micro vision tasks whose training time is negligible, yet together they probe a broad spectrum of representational capacities. Built upon these tasks, we introduce a budget-adaptive performance evaluator to produce the most accurate ranking attainable within the limit. Experiments on popular NAS benchmarks show that our method achieves a ranking correlation higher than existing methods. Furthermore, we construct a search space from prevalent neural blocks and run our method at a cost close to training-free methods; the discovered architecture surpasses the current state-of-the-art under identical training recipes. Our codes are available at https://github.com/fanyi-plus/tf-nas.

Continual Learning (CL) enables Vision-Language Models (VLMs) to acquire new capabilities while retaining prior knowledge, for example, by employing task-specific adapters. Existing CL approaches typically optimize these adapters to convergence, often with (near-)orthogonality constraints to reduce interference; however, isolating adapters in orthogonal subspaces can suppress cross-task transfer and sharing. To address this problem, we provide a new perspective based on PAC-Bayesian analysis: once the per-task optimization has converged, adapters should be further shaped to satisfy \underline P hase-like tr\underline A nsition \underline C ons\underline T raints (PACT) -- a two-part formulation that (i) specifies a phase-like transition relation among adapters and (ii) imposes explicit constraints that enforce this relation. Under PACT, adapter dynamics resemble the phase transition of water: the system gravitates toward either a "frozen" (history-preserving, tightly constrained) or a "melted" (task-adaptive, free) regime, while moving between them smoothly rather than via hard thresholds. We operationalize PACT by coupling stability and plasticity regularizers within a two-branch Vision Transformer (ViT), seeding adapters with a Stable Adapter Initialization (SAI), and introducing a Prior Anchoring (PA) mechanism, thereby inducing phase-like adapter dynamics. Across diverse CL settings, PACT surpasses state-of-the-art methods while reducing the number of trainable parameters by 36.96% relative to standard adapter-based baselines. Our code will be released publicly.

While Vision-Language Models (VLMs) have achieved remarkable performance across diverse downstream tasks, recent studies have shown that they can inherit social biases from the training data and further propagate them into downstream applications. To address this issue, various debiasing approaches have been proposed, yet most of them aim to improve fairness without having a theoretical guarantee that the utility of the model is preserved. In this paper, we introduce a debiasing method that yields a closed-form solution in the cross-modal space, achieving Pareto-optimal fairness with bounded utility losses. Our method is training-free, requires no annotated data, and can jointly debias both visual and textual modalities across downstream tasks. Extensive experiments show that our method outperforms existing methods in debiasing VLMs across diverse fairness metrics and datasets for both group and intersectional fairness in downstream tasks such as zero-shot image classification, text-to-image retrieval, and text-to-image generation while preserving task performance. Code will be made available upon acceptance.

Computer vision and robotics applications ranging from augmented reality to robot autonomy in large-scale environments require spatio-temporal memory frameworks that capture both geometric structure for accurate language-grounding and semantic detail. Existing methods face a tradeoff, where producing rich open-vocabulary descriptions comes at the expense of real-time performance when these descriptions have to be grounded in 3D. To address these challenges, we propose Describe Anything, Anywhere, at Any Moment (DAAAM), a novel spatio-temporal memory framework for large-scale and real-time 4D scene understanding. DAAAM introduces a novel optimization-based frontend to infer detailed semantic descriptions from localized captioning models, such as the Describe Anything Model (DAM), using batch processing to speed up inference by an order of magnitude for online deployment. It leverages such semantic understanding to build a hierarchical 4D scene graph (SG), which acts as an effective globally spatially and temporally consistent memory representation. DAAAM constructs 4D SGs with detailed, geometrically grounded descriptions while maintaining real-time performance. We show that DAAAM's 4D SG interfaces well with a tool-calling agent for inference and reasoning. We thoroughly evaluate DAAAM in the complex task of spatio-temporal question answering (SQA) on the NaVQA benchmark and show its generalization capabilities for sequential task grounding on the SG3D benchmark. We further curate an extended OC-NaVQA benchmark for large-scale and long-time evaluations. DAAAM achieves state-of-the-art results in both tasks, improving OC-NaVQA question accuracy by 53.6%, reducing position errors by 21.9% and temporal errors by 21.6%, and improving SG3D task grounding accuracy by 27.8% over the most competitive baselines. We release our data and code open-source.

Recent advances in video generation models have achieved impressive results. However, these models heavily rely on the use of high-quality data that combines both high visual quality and high motion quality. In this paper, we identify a key challenge in video data curation: the Motion-Vision Quality Dilemma. We discovered that visual quality and motion intensity inherently exhibit a negative correlation, making it hard to obtain golden data that excels in both aspects. To address this challenge, we first examine the hierarchical learning dynamics of video diffusion models and conduct gradient-based analysis on quality-degraded samples. We discover that quality-imbalanced data can produce gradients similar to golden data at appropriate timesteps. Based on this, we introduce the novel concept of Timestep selection in Training Process. We propose Timestep-aware Quality Decoupling (TQD), which modifies the data sampling distribution to better match the model's learning process. For certain types of data, the sampling distribution is skewed toward higher timesteps for motion-rich data, while high visual quality data is more likely to be sampled during lower timesteps. Through extensive experiments, we demonstrate that TQD enables training exclusively on separated imbalanced data to achieve performance surpassing conventional training with better data, challenging the necessity of perfect data in video generation. Moreover, our method also boosts model performance when trained on high-quality data, showcasing its effectiveness across different data scenarios.

AI-assisted coding has rapidly reshaped software practice and research workflows, yet today's models still struggle to produce correct code for complex 3D geometric vision. If models could reliably write such code, the research of our community would change substantially. To measure progress toward that goal, we introduce GeoCodeBench, a PhD-level benchmark that evaluates coding for 3D vision. Each problem is a fill-in-the-function implementation task curated from representative papers at recent venues: we first let a tool propose candidate functions from official repositories, then perform careful human screening to select core 3D geometric components. For every target, we generate diverse, edge-case unit tests, enabling fully automatic, reproducible scoring. We evaluate eight representative open- and closed-source models to reflect the current ecosystem. The best model, GPT-5, attains only 36.6% pass rate, revealing a large gap between current capabilities and dependable 3D scientific coding. GeoCodeBench organizes tasks into a two-level hierarchy: General 3D capability (geometric transformations and mechanics/optics formulation) and Research capability (novel algorithm implementation and geometric logic routing). Scores are positively correlated across these axes, but research-oriented tasks are markedly harder. Context ablations further show that "more paper text" is not always better: cutting off at the Method section statistically outperforms full-paper inputs, highlighting unresolved challenges in long-context scientific comprehension. Together, these findings position GeoCodeBench as a rigorous testbed for advancing from generic coding to trustworthy 3D geometric vision coding.

Large vision-language models (VLMs) are highly vulnerable to multimodal jailbreak attacks that exploit visual-textual interactions to bypass safety guardrails. In this paper, we present DTR, a novel inference-time defense that mitigates multimodal jailbreak attacks through optimizing the model's key-value (KV) caches. Rather than relying on curated safety-specific data or costly image-to-text conversion, we introduce a new formulation of the safety-relevant distributional shift induced by the visual modality. This formulation enables DTR to dynamically adjust visual token weights, minimizing the impact of adversarial visual inputs while preserving the model's general capabilities and inference efficiency. Extensive evaluation across diverse VLMs and attack benchmarks demonstrates that DTR outperforms existing defenses in both attack robustness and benign-task performance, marking the first successful application of KV cache optimization for safety enhancement in multimodal foundation models. The code for replicating DTR is available at: https://github.com/TanqiuJiang/DTR.

Multi-Label Recognition (MLR) based on Vision-Language Models (VLMs) aims to leverage their pre-trained knowledge to better adapt complex recognition scenarios, thereby enhancing model robustness. However, for realistic decentralized applications requiring federated learning, adapting VLMs to each client that possesses private and heterogeneous data can cause the model to overfit spurious label correlations, consequently triggering irrelevant categories when encountering new samples. To tackle this problem, we reconsider the federated learning for MLR with a causal model, in which we adopt a front-door adjustment and decouple the MLR modeling process by intermediate variables that magnify the oracle label co-occurrence. Guided by our analysis, we propose our FedMPT, the first method specifically designed for federated MLR. The core idea of FedMPT is to leverage generalizable conditions to steer federated MLR to mitigate erroneous label activations. To achieve this, FedMPT introduces an Large Language Model (LLM)-driven pipeline to decipher the underlying conditions that govern label dependencies. Furthermore, we introduce an optimal transport between the condition-enriched prompts and the image patches to uncover multiple region-level semantics. Finally, we generate synergistic predictions from different conditions with a crafted gating mechanism. Experiments on multiple benchmark datasets show that our proposed approach achieves competitive results and outperforms SOTA methods under varied settings.

Vision-language models (VLMs) struggle with 3D-related tasks such as spatial cognition and physical understanding, which are crucial for real-world applications like robotics and embodied agents. We attribute this to a modality gap between the 3D tasks and the 2D training of VLM, which led to inefficient retrieval of 3D information from 2D input. To bridge this gap, we introduce SandboxVLM, a simple yet effective framework that leverages abstract bounding boxes to encode geometric structure for VLMs. Specifically, we design a 3D Sandbox reconstruction and perception pipeline comprising four stages: generating multi-view priors with abstract control, proxy elevation, multi-view voting and clustering, and 3D-aware reasoning. Evaluated in zero-shot settings across multiple benchmarks and VLM backbones, our approach consistently improves spatial intelligence, achieving an 8.3% gain on SAT Real compared with baseline methods for instance. These results demonstrate that presenting VLMs with 3D abstractions substantially enhances their 3D reasoning ability without additional training, suggesting new possibilities for general-purpose embodied intelligence.

Multimodal Large Language Models (MLLMs) increasingly need to forget specific knowledge, such as unsafe or private information, without full retraining. However, existing unlearning methods often disrupt vision-language alignment, causing models to reject both harmful and benign queries simultaneously. We trace this failure to the projector network: during unlearning, its Jacobian becomes severely ill-conditioned, leading to unstable optimization and drift in cross-modal embeddings. We introduce SineProject, a simple approach that augments the frozen projector with sinusoidally modulated trainable parameters that improve the Jacobian's spectral conditioning and stabilize alignment throughout unlearning. Evaluated across standard safety and privacy unlearning benchmarks using LLaVA-v1.5-7B and 13B, SineProject reduces benign-query refusals while achieving complete forgetting of targeted information, delivering state-of-the-art forget-retain trade-offs with negligible computational overhead

Human perception of visual similarity is inherently adaptive and subjective, depending on the users' interests and focus. However, most image retrieval systems fail to reflect this flexibility, relying on a fixed, monolithic metric that cannot incorporate multiple conditions simultaneously. To address this, we propose CLAY, an adaptive similarity computation method that reframes the embedding space of pretrained Vision-Language Models (VLMs) as a text-conditional similarity space without additional training. This design separates the textual conditioning process and visual feature extraction, allowing highly efficient and multi-conditioned retrieval with fixed visual embeddings. We also construct a synthetic evaluation dataset CLAY-EVAL, for comprehensive assessment under diverse conditioned retrieval settings. Experiments on standard datasets and our proposed dataset show that CLAY achieves state-of-the art retrieval accuracy and notable computational efficiency compared to previous works.