CVPR 2026 Papers

Streaming video reasoning requires models to operate in a setting where history grows without bound while meaningful evidence remains scarce. In such a landscape, relevant signal is like an oasis -- small, critical, and easily lost in a desert of redundancy. Enlarging memory only widens the desert; aggressive compression dries up the oasis. The real difficulty lies in discovering where to look, not how much to remember. We therefore introduce OASIS, a novel framework for streaming video reasoning that tackles this challenge through structured, on-demand retrieval. It organizes streaming history into hierarchical events and performs reasoning as controlled refinement -- short-context inference first, followed by semantically grounded retrieval only when uncertainty arises. As the retrieval is driven by high-level intent rather than embedding similarity, the retrieved memory is substantially more accurate and less noisy. Additionally, the mechanism is plug-and-play, training-free, and readily attaches to different streaming MLLM backbones. Experiments across multiple benchmarks and backbones show that OASIS achieves strong gains in long-horizon accuracy and compositional reasoning with bounded token cost and low request delay.

Metric Cross-View Geo-Localization (MCVGL) aims to estimate the 3-DoF camera pose (position and heading) by matching ground and satellite images. In this work, instead of pinhole and satellite images, we study robust MCVGL using holistic panoramas and OpenStreetMap (OSM). To this end, we establish a large-scale MCVGL benchmark dataset, CV-RHO, with over 2.7M images under different weather and lighting conditions, as well as sensor noise. Furthermore, we propose a model termed RHO with a two-branch Pin-Pan architecture for accurate visual localization. A Split-Undistort-Merge (SUM) module is introduced to address the panoramic distortion, and a Position-Orientation Fusion (POF) mechanism is designed to enhance the localization accuracy. Extensive experiments prove the value of our CV-RHO dataset and the effectiveness of the RHO model, with a significant performance gain up to 20% compared with the state-of-the-art baselines. Project page: https://github.com/InSAI-Lab/RHO.

Thermal cameras provide reliable visibility in darkness and adverse conditions, but thermal imagery remains significantly harder to use for novel view synthesis (NVS) than visible-light images. This difficulty stems primarily from two characteristics of affordable thermal sensors. First, thermal images have extremely low dynamic range, which weakens appearance cues and limits the gradients available for optimization. Second, thermal data exhibit rapid frame-to-frame photometric fluctuations together with slow radiometric drift, both of which destabilize correspondence estimation and create high-frequency floater artifacts during view synthesis, particularly when no RGB guidance (beyond camera pose) is available. Guided by these observations, we introduce a lightweight preprocessing and splatting pipeline that expands usable dynamic range and stabilizes per-frame photometry. Our approach achieves state-of-the-art performance across thermal-only NVS benchmarks, without requiring any dataset-specific tuning.

Classifier-Free Guidance (CFG) is a cornerstone of modern conditional diffusion models, yet its reliance on the fixed or heuristic dynamic guidance weight is predominantly empirical and overlooks the inherent dynamics of the diffusion process. In this paper, we provide a rigorous theoretical analysis of the Classifier-Free Guidance. Specifically, we establish strict upper bounds on the score discrepancy between conditional and unconditional distributions at different timesteps based on the diffusion process.This finding explains the limitations of fixed-weight strategies and establishes a principled foundation for time-dependent guidance. Motivated by this insight, we introduce **Control Classifier-Free Guidance (C^2FG)**, a novel, training-free, and plug-in method that aligns the guidance strength with the diffusion dynamics via an exponential decay control function. Extensive experiments demonstrate that C^2FG is effective and broadly applicable across diverse generative tasks, while also exhibiting orthogonality to existing strategies.

Methods based on diffusion models (DMs) for solving inverse problems (IPs) have recently achieved remarkable performance. However, DM-based methods typically struggle against outliers, which are common in real-world measurements. In this work, to tackle IPs with outliers, we first refine the measurement via explicit noise estimation to mitigate the effect of noise. Subsequently, we formulate an iteratively reweighted least squares objective based on the Huber loss to address the outliers. We propose a method utilizing gradient descent to approximately solve the corresponding optimization problem for the robust objective. To avoid delicate tuning of the learning rate required by the gradient descent method, we further employ the conjugate gradient method with an efficient strategy for updating. Extensive experiments on multiple image datasets for linear and nonlinear tasks under various conditions demonstrate that our proposed methods exhibit robustness to outliers and outperform recent DM-based methods in most cases.

Lookup-free quantization has received much attention due to its efficiency on parameters and scalability to a large codebook. In this paper, we present a unified formulation of different non-parametric quantization methods through the lens of lattice coding. The geometry of lattice codes explains the necessity of auxiliary loss terms when training auto-encoders with certain existing lookup-free quantization variants such as BSQ. As a step forward, we explore a few possible candidates, including random lattices, generalized Fibonacci lattices, and densest sphere packing lattices. Among all, we find the Leech lattice-based quantization method, which is dubbed as Spherical Leech Quantization (\Lambda_ 24 -SQ), leads to both a simplified training recipe and an improved reconstruction-compression tradeoff thanks to its high symmetry and even distribution on the hypersphere. In image tokenization and compression tasks, this quantization approach achieves better reconstruction quality across all metrics than BSQ, the best prior art, while consuming slightly fewer bits. The improvement also extends to state-of-the-art auto-regressive image generation frameworks.

Conditional image generation methods are increasingly used in human-centric applications, yet existing human amodal completion (HAC) models offer users limited control over the completed content. Given an occluded person image, they hallucinate invisible regions while preserving visible ones, but cannot reliably incorporate user-specified constraints such as a desired pose or spatial extent. As a result, users often resort to repeatedly sampling the model to obtain a satisfactory output. Pose-guided person image synthesis (PGPIS) methods allow explicit pose conditioning, but frequently fail to preserve the instance-specific visible appearance and tend to be biased toward the training distribution, even when built on strong diffusion model priors. To address these limitations, we introduce promptable human amodal completion (PHAC), a new task that completes occluded human images while satisfying both visible appearance constraints and multiple user prompts. Users provide simple point-based prompts, such as additional joints for the target pose or bounding boxes for desired regions; these prompts are encoded using ControlNet modules specialized for each prompt type. These modules inject the prompt signals into a pre-trained diffusion model, and we fine-tune only the cross-attention blocks to obtain strong prompt alignment without degrading the underlying generative prior. To further preserve visible content, we propose an inpainting-based refinement module that starts from a slightly noised coarse completion, faithfully preserves the visible regions, and ensures seamless blending at occlusion boundaries. Extensive experiments on standard HAC and PGPIS benchmarks show that our approach produces more physically plausible, higher-quality completions with significantly improved prompt alignment compared to existing amodal completion and pose-guided synthesis methods.

Humans inhabit a physical 4D world, where spatial geometry and semantic content evolve over time, forming a dynamic reality. While current Multimodal Large Language Models (MLLMs) demonstrate strong capabilities in understanding static visual inputs, it remains unclear whether they can effectively "think in dynamics," i.e., perceive, track, and reason about spatio-temporal evolution in complex scenes.To systematically evaluate these abilities, we introduce \texttt Dyn-Bench , a large-scale benchmark designed to assess spatio-temporal reasoning and localized dynamics perception. Constructed through multi-stage filtering over massive 2D and 4D data sources, \texttt Dyn-Bench provides a high-quality collection of diverse dynamic scenes, consisting of 1k videos, 7k visual question answering (VQA) pairs, and 3k dynamic object grounding samples.We comprehensively study general-purpose, spatial-aware, and region-level MLLMs to understand how they "think in dynamics" from both linguistic and visual perspectives. Our results reveal that existing models struggle to jointly excel in both spatio-temporal reasoning and dynamic object grounding, often producing inconsistent interpretations of motion and interaction. Conventional prompting strategies i.e., chain-of-thought or caption-based hints) provide only limited improvements.In contrast, structured integration approaches, including Mask-Guided Fusion and the Spatio-Temporal Textual Cognitive Map (ST-TCM), substantially enhance MLLMs' dynamic perception and spatio-temporal reasoning in an evolving 4D world. These findings underscore the importance of explicit spatio-temporal structural cues to bridge the gap between static perception and dynamic reasoning in MLLMs.

Single-image reflection removal (SIRR) is a highly ill-posed and computationally demanding problem. Existing CNN or Transformer-based methods often rely on large receptive fields and heavy computation, limiting their deployment on resource-constrained devices. To address this, we propose LightRR, a lightweight yet effective reflection removal network that unifies a wavelet-based mechanism and State Space Modeling (SSM).Specifically, we introduce an Asymmetric Frequency Mamba Block (AFM), which leverages the Discrete Wavelet Transform (DWT) to decompose features into low- and high-frequency components. This allows for targeted modeling of frequency-specific dependencies via Mamba-based state space dynamics.This design not only captures long-range context efficiently but also reduces spatial resolution and computation while preserving critical details.Furthermore, a knowledge distillation-enhanced encoder allows the network to inherit the representational power of large pre-trained models during training, enabling lightweight inference.Extensive experiments on multiple real-world benchmarks demonstrate that LightRR achieves performance comparable to state-of-the-art methods, while using only 3.01% of the parameters and 5.22% of the FLOPs (vs. RDNet), highlighting its superior balance between accuracy and efficiency.

Humans intuitively move to sound, but current humanoid robots lack expressive improvisational capabilities, confined to predefined motions or sparse commands. Generating motion from audio and then retargeting it to robots relies on explicit motion reconstruction, leading to cascaded errors, high latency, and disjointed acoustic-actuation mapping. We propose RoboPerform, the first unified audio-to-locomotion framework that can directly generate music-driven dance and speech-driven co-speech gestures from audio. Guided by the core principle of "motion = content + style", the framework treats audio as implicit style signals and eliminates the need for explicit motion reconstruction. RoboPerform integrates a ResMoE teacher policy for adapting to diverse motion patterns and a diffusion-based student policy for audio style injection. This retargeting-free design ensures low latency and high fidelity. Experimental validation shows that RoboPerform achieves promising results in physical plausibility and audio alignment, successfully transforming robots into responsive freestyle performers capable of reacting to audio.

Language is a natural way to command robots, but converting a single instruction into a long-horizon, contact-rich hand-object interaction remains challenging: synthesized references are noisy, human-to-robot retargeting introduces embodiment bias, and fixed-reference tracking lets small errors snowball. We address this with AdaDexTrack, a modulator-in-the-loop framework for language-conditioned manipulation tracking. A distilled generalist tracker serves as the skill carrier, while a tightly aligned modulator performs three feedback corrections: reference modulation (continual adjustment of what to track), object-latent modulation (online adaptation of the object representation to recruit suitable skills), and positional-target modulation (small state-dependent refinements for execution). The tracker is learned via large-scale specialist to generalist distillation on a corpus of language-conditioned hand-object trajectories; the modulator is trained with RL under the same task objective, ensuring tight coupling. Across large-scale evaluations, AdaDexTrack consistently outperforms prior SOTA on unseen-trajectory and unseen-object sets in both average tracking error and success rate, demonstrating robustness and generalization. We further show zero-shot sim-to-real transfer on real hardware, where adding the modulator yields substantial gains over a tracker-only variant. AdaDexTrack reframes language-conditioned dexterous manipulation as modulated tracking, replacing the open-loop, fixed-reference tracking with in-loop modulation that adjusts the reference, object latent, and positional target, yielding drift-resistant execution from noisy text references.

Open-vocabulary semantic segmentation extends pixel-level recognition to arbitrary text-described categories. Despite strong global semantic understanding, vision-language models such as CLIP exhibit limited spatial precision and semantic ambiguity across large vocabularies, constraining their effectiveness for dense prediction. We present S2C2Seg, a training-free framework that integrates with existing methods through Category Subset Selection (CSS) and Consistent Semantic Guidance (CSG). CSS employs three complementary scoring functions to filter category candidates: CLIP-based global semantic similarity, spatial presence from dense prediction models, and multi-view consistency via alignment and conditional entropy. This joint exploitation of semantic, spatial, and consistency cues reduces category redundancy and semantic ambiguity. CSG adaptively fuses CLIP global features with local spatial predictions through category-specific confidence weighting, applying stronger regularization to high-similarity categories for correcting prediction biases while preserving spatial precision for low-confidence categories. Extensive experiments across eight benchmarks demonstrate broad applicability: when integrated with SCLIP, ProxyCLIP, and CorrCLIP, S2C2Seg achieves consistent improvements of 3.4 to 9.7 percentage points in mIoU, establishing a new state-of-the-art of 51.2% average mIoU.

State-Space Models (SSMs) excel at capturing long-range dependencies with structured recurrence, making them well-suited for sequence modeling. However, their evolving internal states pose unique challenges in Continual Learning (CL). Without access to the full distribution of previous tasks, updates to the state-space dynamics become unconstrained, leading to catastrophic forgetting. To address this, we propose Inf-SSM, a geometry-aware regularization framework for CL in SSMs. It constrains state evolution via the infinite-dimensional Grassmannian of SSM observability subspaces, without requiring any exemplars from past tasks. Unlike classical CL methods that restrict weight updates, Inf-SSM directly regularizes the infinite-horizon state evolution encoded by the extended observability subspace of the SSM. We show that enforcing this regularization requires solving a matrix equation known as the Sylvester equation, which typically incurs \mathcal O (n^3) complexity. Thus, we develop a \mathcal O (n^2) solution by exploiting the structure and properties of SSMs. This leads to an efficient regularization mechanism that can be seamlessly integrated into existing CL methods. Comprehensive experiments on challenging benchmarks of ImageNet-R, CIFAR-100, and Caltech-256 demonstrate a significant reduction in forgetting while improving accuracy across sequential tasks.

Decomposing a scene into its reflectance and shading is a challenge due to the lack of extensive ground-truth data for real-world scenes. We introduce a novel physics-based approach for intrinsic image decomposition using a pair of visible and thermal images. We leverage the principle that light not reflected from an opaque surface is absorbed and detected as heat by a thermal camera. This allows us to relate the ordinalities (or relative magnitudes) between visible and thermal image intensities to the ordinalities of shading and reflectance. The ordinalities enable dense self-supervision of an optimizing neural network to recover shading and reflectance. We perform quantitative evaluations with known reflectance and shading under natural and artificial lighting, and qualitative experiments across diverse scenes. The results demonstrate superior performance over both physics-based and recent learning-based methods, providing a path toward scalable real-world data curation with supervision.

We study concept-level forgetting in pretrained vision models: removing an entire semantic category so the system no longer recognizes that object in unseen images and contexts, rather than merely forgetting specific training examples. Prior work either applies blunt global projections or fine-tunes parameters, which can introduce collateral damage to unrelated features, add compute, and become unstable as forgetting strength increases. We introduce Contrastive Subnet Erasure (CSE), a training-free, encoder-centric edit that targets a compact set of channels most responsible for the class and attenuates them in a calibrated manner. The modification is algebraically folded into the subsequent layer, yielding no inference-time overhead and leaving task heads unchanged. To evaluate whether forgetting generalizes beyond the data used to specify the class, we introduce a cross dataset protocol in which the class is defined on a source dataset and performance is measured on a disjoint target dataset drawn from a different distribution with no shared images. This setup tests whether the model still fails to recognize the object when it looks different or appears in new scenes, and it helps avoid overfitting to patterns in the source dataset. Across CIFAR 10, CIFAR 100, and ImageNet under this protocol, CSE achieves stronger forgetting of the target class while better preserving non target utility than existing baselines in both single class and multi class settings. Overall, CSE provides a simple, stable, and deployment-ready mechanism for class-level unlearning in vision.

Deep Neural Networks (DNNs) have revolutionized numerous industries, yet their decision-making processes remain largely opaque. Most existing explanation methods visualize the importance of image regions that influence a classifier's decisions, but they predominantly focus on identifying regions with positive contributions, often overlooking those with negative impacts. In this paper, we introduce a novel black-box explanation method, the Metropolis-Hastings Explainer (MHE), designed to provide confidence-faithful explanations. MHE enhances the fidelity of explanations by ensuring that the explained regions closely align with the original confidence score, sampling instances that best match the classifier's confidence. Furthermore, MHE improves sampling efficiency by utilizing existing valid samples to explore more potential valid ones, reducing computational overhead. To enhance the clarity of explanations, MHE prioritizes valid samples with smaller areas when other factors are equal, thereby reducing the explanation area. Building upon the MHE framework, we propose two extensions: MHE-e, which focuses exclusively on regions with positive contributions, and MHE-pro, which refines explanation quality by integrating multi-scale information. MHE-pro progressively regions, optimizing both sampling efficiency and explanation quality. Experimental results demonstrate that MHE delivers superior and stable explanation quality across various models, including ResNet50, VGG16, ViT, DINO, and CLIP, on datasets such as ImageNet, CUB-200-2011, and VOC2012, providing explanations that closely approximate the original classification confidence.

We present Pressure2Motion, a novel motion capture algorithm that reconstructs human motion from a ground pressure sequence and text prompt. At inference time, Pressure2Motion requires only a pressure mat, eliminating the need for specialized lighting setups, cameras, or wearable devices, making it suitable for privacy-preserving, low-light, and low-cost motion capture scenarios. Such a task is severely ill-posed due to the indeterminacy of pressure signals with respect to full-body motion. To address this issue, we introduce Pressure2Motion, a generative model that leverages pressure features as input and utilizes a text prompt as a high-level guiding constraint to resolve ambiguities. Specifically, our model adopts a dual-level feature extractor to accurately interpret pressure data, followed by a hierarchical diffusion model that discerns broad-scale movement trajectories and subtle posture adjustments. Both the physical cues gained from the pressure sequence and the semantic guidance derived from descriptive texts are leveraged to guide the motion estimation with precision. To the best of our knowledge, Pressure2Motion is a pioneering work in leveraging both pressure data and linguistic priors for motion reconstruction, and the established MPL benchmark is the first benchmark for this novel motion capture task. Experiments show that our method generates high-fidelity, physically plausible motions, establishing a new state of the art for this task. The codes and benchmarks will be publicly released upon publication.

Vision-Language Models (VLMs) often struggle with robust 3D spatial reasoning. Prevailing methods that rely on fine-tuning with 3D visual question-answering (VQA) datasets may overfit dataset-specific biases, while integrating specialized 3D visual encoders is often inflexible and cumbersome. In this paper, we argue that genuine spatial understanding should emerge from learning fundamental geometric priors, not only from high-level VQA supervision. We propose GASP (Geometric-Aware Spatial Priors), a framework that injects these priors directly into the LLM's transformer layers. GASP employs a small correspondence head, applied as a deep supervision signal across all layers, and is trained with a dual objective leveraging ground-truth geometry from large-scale video scenes: a contrastive loss on ground-truth point correspondences enforces 2D view-invariance, while a depth consistency supervision resolves 3D geometric ambiguities. Our analysis first provides a diagnostic showing that standard VLMs' internal correspondence matching accuracy is very low (often below 5%). We then demonstrate that our training substantially improves this behavior, boosting peak layer-wise correspondence to over 70% and maintaining over 85% temporal robustness while baselines remain below 5%. These internal improvements translate to significant gains on downstream spatial benchmarks---including +18.2% on All-Angles Bench and +29.0% on VSI-Bench---all without training on any 3D VQA data. Our findings indicate that learning from fundamental geometric priors is a promising and generalizable pathway towards VLMs with more reliable 3D spatial reasoning.

3D Referring Expression Segmentation (3D-RES) aims to segment objects in point clouds according to language descriptions. Unlike common practices in 2D that utilize learnable query embeddings, recent 3D-RES methods typically generate queries directly from 3D points. However, this direct coupling of queries to raw point clouds introduces new challenges: an impractically large number of queries derived from massive point cloud data and a reliance on non-deterministic sampling algorithms. In this paper, we propose a Semantic-based Adaptive Query Network (SAQN), which introduces a novel query strategy for 3D-RES. Instead of generating queries from points, SAQN employs a learnable query vector for each semantic class. This approach drastically reduces the number of queries while maintaining the advantage of avoiding Hungarian matching through implicit class alignment. Additionally, to address potential cross-object ambiguity within semantic classes, we introduce supplementary queries that are adaptively fused with each class query to disambiguate and enrich representations. Comprehensive experiments show that SAQN achieves state-of-the-art performance while reducing the number of queries.