ICCV 2025 论文列表

Recent interactive matting methods have demonstrated satisfactory performance in capturing the primary regions of objects, but they fall short in extracting fine-grained details in edge regions. Diffusion models trained on billions of image-text pairs, demonstrate exceptional capability in modeling highly complex data distributions and synthesizing realistic texture details, while exhibiting robust text-driven interaction capabilities, making them an attractive solution for interactive matting. To this end, we propose SDMatte, a diffusion-driven interactive matting model, with three key contributions. First, we exploit the powerful priors of the pre-trained U-Net within diffusion models and transform the text-driven interaction mechanism into a visual prompt-driven interaction mechanism to enable interactive matting. Second, we integrate coordinate embeddings of visual prompts and opacity embeddings of objects into U-Net, enhancing SDMatte's sensitivity to spatial position information and opacity information. Third, we propose a masked self-attention mechanism and a visual prompt-driven interaction mechanism that enable the model to focus on areas specified by visual prompts, leading to better performance. Extensive experiments on multiple datasets demonstrate the superior performance of our method, validating its effectiveness in interactive matting. Code will be made publicly available.

Vision-based semantic scene completion (SSC) is able to predict complex scene information from limited 2D images, which has attracted widespread attention. Currently, SSC methods typically construct unified voxel features containing both geometry and semantics, which lead to different depth positions in occluded regions sharing the same 2D semantic information, resulting in ambiguous semantic segmentation. To address this problem, we propose SDFormer, a novel SAM-assisted Dual-channel Voxel Transformer framework for SSC. We uncouple the task based on its multi-objective nature and construct two parallel sub-networks: a semantic constructor (SC) and a geometric refiner (GR). The SC utilizes the Segment Anything Model (SAM) to construct dense semantic voxel features from reliable visible semantic information in the image. The GR accurately predicts depth positions and then further adjusts the semantic output by SAM. Additionally, we design a Semantic Calibration Affinity to enhance semantic-aware transformations in SC. Within the GR, Shape Segments Interactive and Learnable mask generation module to emphasize the spatial location of semantics to obtain fine-grained voxel information. Extensive qualitative and quantitative results on the SemanticKITTI and SSCBench-KITTI-360 datasets show that our method outperforms state-of-the-art approaches.

Recovering 3D object pose and shape from a single image is a challenging and ill-posed problem. This is due to strong (self-)occlusions, depth ambiguities, the vast intra- and inter-class shape variance, and the lack of 3D ground truth for natural images. Existing deep-network methods are trained on synthetic datasets to predict 3D shapes, so they often struggle generalizing to real-world images. Moreover, they lack an explicit feedback loop for refining noisy estimates, and primarily focus on geometry without directly considering pixel alignment. To tackle these limitations, we develop a novel render-and-compare optimization framework, called SDFit. This has three key innovations: First, it uses a learned category-specific and morphable signed-distance-function (mSDF) model, and fits this to an image by iteratively refining both 3D pose and shape. The mSDF robustifies inference by constraining the search on the manifold of valid shapes, while allowing for arbitrary shape topologies. Second, SDFit retrieves an initial 3D shape that likely matches the image, by exploiting foundational models for efficient look-up into 3D shape databases. Third, SDFit initializes pose by establishing rich 2D-3D correspondences between the image and the mSDF through foundational features. We evaluate SDFit on three image datasets, i.e., Pix3D, Pascal3D+, and COMIC. SDFit performs on par with SotA feed-forward networks for unoccluded images and common poses, but is uniquely robust to occlusions and uncommon poses. Moreover, it requires no retraining for unseen images. Thus, SDFit contributes new insights for generalizing in the wild. Code is available at https://anticdimi.github.io/sdfit.

Language-conditioned robot manipulation in the continuous spectrum presents a persistent challenge due to the difficult of mapping states to target actions. Previous methods face limitations in effectively modeling object states, primarily due to their reliance on executing ambiguous instructions devoid of explicit state information. In response, we present SD^2Actor, a zero-shot robotic manipulation framework that possesses the capability to generate precise actions in continuous states. Specifically, given the novel instructions, we aim to generate instruction-following and accurate robot manipulation actions. Instead of time-consuming optimization and finetuning, our zero-shot method generalizes to any object state with a wide range of translations and versatile rotations. At its core, we quantify multiple base states in the training set and utilize their combination to refine the target action generated by the diffusion model. To obtain novel state representations, we initially employ LLMs to extract the novel state from the instruction and decompose it into multiple learned base states. We then employ the linear combination of base state embeddings to produce novel state features. Moreover, we introduce the orthogonalization loss to constrain the state embedding space, which ensures the validity of linear interpolation. Experiments demonstrate that SD^2Actor outperforms state-of-the-art methods across a diverse range of manipulation tasks in ARNOLD Benchmark. Moreover, SD^2Actor can effectively learn generalizable policies from a limited number of human demonstrations, achieving promising accuracy in a variety of real-world manipulation tasks.

Most existing remote sensing instance segmentation approaches are designed for close-vocabulary prediction, limiting their ability to recognize novel categories or generalize across datasets. This restricts their applicability in diverse Earth observation scenarios. To address this, we introduce open-vocabulary (OV) learning for remote sensing instance segmentation. While current OV segmentation models perform well on natural image datasets, their direct application to remote sensing faces challenges such as diverse landscapes, seasonal variations, and the presence of small or ambiguous objects in aerial imagery. To overcome these challenges, we propose SCORE (Scene Context matters in Open-vocabulary REmote sensing instance segmentation), a framework that integrates multi-granularity scene context, i.e., regional context and global context, to enhance both visual and textual representations. Specifically, we introduce Region-Aware Integration, which refines class embeddings with regional context to improve object distinguishability. Additionally, we propose Global Context Adaptation, which enriches naive text embeddings with remote sensing global context, creating a more adaptable and expressive linguistic latent space for the classifier. We establish new benchmarks for OV remote sensing instance segmentation across diverse datasets. Experimental results demonstrate that, our proposed method achieves SOTA performance, which provides a robust solution for large-scale, real-world geospatial analysis. Our code is available at https://github.com/HuangShiqi128/SCORE.

Explicitly disentangling style and content in vision models remains challenging due to their semantic overlap and the subjectivity of human perception. Existing methods propose separation through generative or discriminative objectives, but they still face the inherent ambiguity of disentangling intertwined concepts. Instead, we ask: Can we bypass explicit disentanglement by learning to merge style and content invertibly, allowing separation to emerge naturally? We propose SCFlow, a flow-matching framework that learns bidirectional mappings between entangled and disentangled representations. Our approach is built upon three key insights: 1) Training solely to merge style and content, a well-defined task, enables invertible disentanglement without explicit supervision; 2) flow matching bridges on arbitrary distributions, avoiding the restrictive Gaussian priors of diffusion models and normalizing flows; and 3) a synthetic dataset of 510,000 samples was curated to simulate disentanglement through systematic style-content pairing. Beyond controllable generation tasks, we demonstrate that SCFlow generalizes to ImageNet-1k and WikiArt in zero-shot settings and achieves competitive performance, highlighting that disentanglement naturally emerges from the invertible merging process. Code and dataset: https://github.com/CompVis/SCFlow

While contrastive pre-training is widely employed, its data efficiency problem has remained relatively under-explored thus far. Existing methods often rely on static coreset selection algorithms to pre-identify important data for training. However, this static nature renders them unable to dynamically track the data utility throughout pre-training, leading to subpar pre-trained models. To address this challenge, our paper introduces a novel dynamic bootstrapping dataset pruning method. It involves pruning data preparation followed by dataset mutation operations, both of which undergo iterative and dynamic updates. We apply this method to two prevalent contrastive pre-training frameworks: CLIP and MoCo, representing vision-language and vision-centric domains, respectively. In particular, we individually pre-train seven CLIP models on two large-scale image-text pair datasets, and two MoCo models on the ImageNet dataset, resulting in a total of 16 pre-trained models. With a data pruning rate of 30-35% across all 16 models, our method exhibits only marginal performance degradation (less than 1% on average) compared to corresponding models trained on the full dataset counterparts across various downstream datasets, and also surpasses several baselines with a large performance margin. Additionally, the byproduct from our method, i.e., coresets derived from the original datasets after pre-training, also demonstrates significant superiority in terms of downstream performance over other static coreset selection approaches. Code is available at https://github.com/guoyang9/SCAN.

In this paper, we introduce SC-Lane, a novel slope-aware and temporally consistent heightmap estimation framework for 3D lane detection. Unlike previous approaches that rely on fixed slope anchors, SC-Lane adaptively determines the fusion of slope-specific height features, improving robustness to diverse road geometries. To achieve this, we propose a Slope-Aware Adaptive Feature module that dynamically predicts the appropriate weights from image cues for integrating multi-slope representations into a unified heightmap. Additionally, a Height Consistency Module enforces temporal coherence, ensuring stable and accurate height estimation across consecutive frames, which is crucial for real-world driving scenarios. To evaluate the effectiveness of SC-Lane, we employ three standardized metrics-Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and threshold-based accuracy-which, although common in surface and depth estimation, have been underutilized for road height assessment. Using the LiDAR-derived heightmap dataset introduced in prior work, we benchmark our method under these metrics, thereby establishing a rigorous standard for future comparisons. Extensive experiments on the OpenLane benchmark demonstrate that SC-Lane significantly improves both height estimation and 3D lane detection, achieving state-of-the-art performance with an F-score of 64.3%, outperforming existing methods by a notable margin. For detailed results and a demonstration video, please refer to our project page: https://parkchaesong.github.io/sclane/

We propose SC-Captioner, a reinforcement learning framework that enables the self-correcting capability of image caption models. Our crucial technique lies in the design of the reward function to incentivize accurate caption corrections. Specifically, the predicted and reference captions are decomposed into object, attribute, and relation sets using scene-graph parsing algorithms. We calculate the set difference between sets of initial and self-corrected captions to identify added and removed elements. These elements are matched against the reference sets to calculate correctness bonuses for accurate refinements and mistake punishments for wrong additions and removals, thereby forming the final reward. For image caption quality assessment, we propose a set of metrics refined from CAPTURE that alleviate its incomplete precision evaluation and inefficient relation matching problems. Furthermore, we collect a fine-grained annotated image caption dataset, RefinedCaps, consisting of 6.5K diverse images from COCO dataset. Experiments show that applying SC-Captioner on large visual-language models can generate better image captions across various scenarios, significantly outperforming the direct preference optimization training strategy.

Unlearning methods for vision-language models (VLMs) have primarily adapted techniques from large language models (LLMs), relying on weight updates that demand extensive annotated forget sets. Moreover, these methods perform unlearning at a coarse granularity, often leading to excessive forgetting and reduced model utility. To address this issue, we introduce SAUCE, a novel method that leverages sparse autoencoders (SAEs) for fine-grained and selective concept unlearning in VLMs. Briefly, SAUCE first trains SAEs to capture high-dimensional, semantically rich sparse features. It then identifies the features most relevant to the target concept for unlearning. During inference, it selectively modifies these features to suppress specific concepts while preserving unrelated information. We evaluate SAUCE on two distinct VLMs, LLaVA-v1.5-7B and LLaMA-3.2-11B-Vision-Instruct, across two types of tasks: concrete concept unlearning (objects and sports scenes) and abstract concept unlearning (emotions, colors, and materials), encompassing a total of 60 concepts. Extensive experiments demonstrate that SAUCE outperforms state-of-the-art methods by 18.04% in unlearning quality while maintaining comparable model utility. Furthermore, we investigate SAUCE's robustness against widely used adversarial attacks, its transferability across models, and its scalability in handling multiple simultaneous unlearning requests. Our findings establish SAUCE as an effective and scalable solution for selective concept unlearning in VLMs.

The open vocabulary capability of 3D models is increasingly valued, as traditional methods with models trained with fixed categories fail to recognize unseen objects in complex dynamic 3D scenes. In this paper, we propose a simple yet effective approach, SAS, to integrate the open vocabulary capability of multiple 2D models and migrate it to 3D domain. Specifically, we first propose Model Alignment via Text to map different 2D models into the same embedding space using text as a bridge. Then, we propose Annotation-Free Model Capability Construction to explicitly quantify the 2D model's capability of recognizing different categories using diffusion models. Following this, point cloud features from different 2D models are fused with the guide of constructed model capabilities. Finally, the integrated 2D open vocabulary capability is transferred to 3D domain through feature distillation. SAS outperforms previous methods by a large margin across multiple datasets, including ScanNet v2, Matterport3D, and nuScenes, while its generalizability is further validated on downstream tasks, e.g., gaussian segmentation and instance segmentation.

This paper presents SANA-Sprint, an efficient diffusion model for ultra-fast text-to-image (T2I) generation. SANA-Sprint is built on a pre-trained foundation model and augmented with hybrid distillation, dramatically reducing inference steps from 20 to 1-4.We introduce three key innovations: (1) We propose a training-free approach that transforms a pre-trained flow-matching model for continuous-time consistency distillation (sCM), eliminating costly training from scratch and achieving high training efficiency. Our hybrid distillation strategy combines sCM with latent adversarial distillation (LADD): sCM ensures alignment with the teacher model, while LADD enhances single-step generation fidelity. (2) SANA-Sprint is a unified step-adaptive model that achieves high-quality generation in 1-4 steps, eliminating step-specific training and improving efficiency. (3) We integrate ControlNet with SANA-Sprint for real-time interactive image generation, enabling instant visual feedback for user interaction. SANA-Sprint establishes a new Pareto frontier in speed-quality tradeoffs, achieving state-of-the-art performance with 7.59 FID and 0.74 GenEval in just 1 step -- outperforming FLUX-schnell (7.94 FID / 0.71 GenEval) while being 10x faster (0.1s vs 1.1s on H100). It also achieves 0.1s (T2I) and 0.25s (ControlNet) latency for 1024x1024 images on H100, and 0.31s (T2I) on an RTX 4090, showcasing its exceptional efficiency and potential for AI-powered consumer applications (AIPC). Code and pre-trained models will be open-sourced.

The Segment Anything Model (SAM) has demonstrated significant potential in medical image segmentation, yet its performance is limited when only a small amount of labeled data is available, while there are abundance of valuable yet often overlooked hierarchical information inherent in medical data. To address this limitation, we draw inspiration from self-supervised learning and propose SAMora, an innovative framework that captures hierarchical medical knowledge by applying complementary self-supervised learning objectives at the image, patch, and pixel levels. To fully exploit the complementarity of hierarchical knowledge within LoRAs, we introduce HL-Attn, a hierarchical fusion module that integrates multi-scale features while maintaining their distinct characteristics. SAMora is compatible with various SAM variants, including SAM2, SAMed and H-SAM. Experimental results on the Synapse, LA, and PROMISE12 datasets demonstrate that SAMora outperforms existing SAM variants, achieving state-of-the-art performance in both few-shot and fully-supervised settings, while reducing fine-tuning epochs by 90%.

Open-vocabulary action recognition (OVAR) extends recognition systems to identify unseen action categories. While large-scale vision-language models (VLMs) like CLIP have enabled OVAR in image domains, their adaptation to event data remains underexplored. Event cameras offer high temporal resolution and inherent privacy preservation, making them suitable for capturing fine-grained motion dynamics. However, leveraging event data for OVAR presents challenges: 1) bridging the domain gap between static image-based models and event streams, and 2) preserving the generalization capabilities of pretrained VLMs in open-vocabulary settings. In this paper, we propose SAMPLE, a lightweight adaptation of VLMs for event-based action recognition, balancing supervised and open-vocabulary performance. We introduce a Temporal-Adaptive Multimodal Prompt Learning strategy that can be divided into: 1) Unimodal prompt on both the event and text branches to learn the data distribution 2) Event-Text cross-modal prompt for representation space alignment 3) Temporal-Adaptive prompt to model temporal dependencies across event data. Extensive evaluations demonstrate that SAMPLE outperforms prior methods across fully supervised, few-shot, base-to-novel and zero-shot settings. Notably, in zero-shot scenarios, SAMPLE achieves gains of +15.46%, +29.76%, and +23.79% on SeAct, DVS128Gesture, and PAF respectively with less commute cost. Our codes are released at https://github.com/JingWang-self/SAMPLE.

Multi-task learning (MTL) enables a joint model to capture commonalities across multiple tasks, reducing computation costs and improving data efficiency. However, a major challenge in MTL optimization is task conflicts, where the task gradients differ in direction or magnitude, limiting model performance compared to single-task counterparts. Sharpness-aware minimization (SAM) minimizes task loss while simultaneously reducing the sharpness of the loss landscape. Our empirical observations show that SAM effectively mitigates task conflicts in MTL. Motivated by these findings, we explore integrating SAM into MTL but face two key challenges. On one hand, both the average loss gradient and individual task gradients--referred to as global and local information--contribute to SAM, but how to combine them remains unclear. On the other hand, directly computing each task gradient introduces significant computational and memory overheads. To address these challenges, we propose SAMO, a lightweight **S**harpness-**A**ware **M**ulti-task **O**ptimization approach, that leverages a joint global-local perturbation. The local perturbations are approximated using only forward passes and are layerwise normalized to improve efficiency. Extensive experiments on a suite of multi-task benchmarks demonstrate both the effectiveness and efficiency of our method.

The academic field of learning instruction-guided visual navigation can be generally categorized into high-level category-specific search and low-level language-guided navigation, depending on the granularity of language instruction, in which the former emphasizes the exploration process, while the latter concentrates on following detailed textual commands. Despite the differing focuses of these tasks, the underlying requirements of interpreting instructions, comprehending the surroundings, and inferring action decisions remain consistent. This paper consolidates diverse navigation tasks into a unified and generic framework -- we investigate the core difficulties of sharing general knowledge and exploiting task-specific capabilities in learning navigation and propose a novel State-Adaptive Mixture of Experts (SAME) model that effectively enables an agent to infer decisions based on different-granularity language and dynamic observations. Powered by SAME, we present a versatile agent capable of addressing seven navigation tasks simultaneously, achieving highly comparable performance to task-specific agents.

We present SAM4D, a multi-modal and temporal foundation model designed for promptable segmentation across camera and LiDAR streams. Unified Multi-modal Positional Encoding (UMPE) is introduced to align camera and LiDAR features in a shared 3D space, enabling seamless cross-modal prompting and interaction. Additionally, we propose Motion-aware Cross-modal Memory Attention (MCMA), which leverages ego-motion compensation to enhance temporal consistency and long-horizon feature retrieval, ensuring robust segmentation across dynamically changing autonomous driving scenes. To avoid annotation bottlenecks, we develop a multi-modal automated data engine that synergizes VFM-driven video masklets, spatiotemporal 4D reconstruction, and cross-modal masklet fusion. This framework generates camera-LiDAR aligned pseudo-labels at a speed orders of magnitude faster than human annotation while preserving VFM-derived semantic fidelity in point cloud representations. We conduct extensive experiments on the constructed Waymo-4DSeg, which demonstrate the powerful cross-modal segmentation ability and great potential in data annotation of proposed SAM4D.

The Segment Anything Model 2 (SAM 2) has emerged as a powerful foundation model for object segmentation in both images and videos. The crucial design of SAM 2 for video segmentation is its memory module, which prompts object-aware memories from previous frames for current frame prediction. However, its greedy-selection memory design suffers from the "error accumulation" problem, where an errored or missed mask will cascade and influence the segmentation of the subsequent frames, which limits the performance of SAM 2 toward complex long-term videos. To this end, we introduce SAM2Long, an improved training-free video object segmentation strategy, which considers the segmentation uncertainty within each frame and chooses the video-level optimal results from multiple segmentation pathways in a constrained tree search manner. In practice, we maintain a fixed number of segmentation pathways throughout the video. For each frame, multiple masks are proposed based on the existing pathways, creating various candidate branches. We then select the same fixed number of branches with higher cumulative scores as the new pathways for the next frame. After processing the final frame, the pathway with the highest cumulative score is chosen as the final segmentation result. Benefiting from its heuristic search design, SAM2Long is robust toward occlusions and object reappearances, and can effectively segment and track objects for complex long-term videos. Without further training, SAM2Long significantly and consistently outperforms SAM 2 on nine VOS benchmarks and three VOT benchmarks. Notably, SAM2Long achieves an average improvement of 3.7 points across all 12 direct comparisons, with gains of up to 5.3 points in J&F on long-term video object segmentation benchmarks such as SA-V and LVOS. The code is released at https://github.com/Mark12Ding/SAM2Long.

While the Segment Anything Model (SAM) transforms interactive segmentation with zero-shot abilities, its inherent vulnerabilities present a single-point risk, potentially leading to the failure of downstream applications. Proactively evaluating these transferable vulnerabilities is thus imperative. Prior adversarial attacks on SAM often present limited transferability due to insufficient exploration of common weakness across domains. To address this, we propose a novel method, Vertex-Refining Simplicial Complex Attack (VeSCA), generating transferable adversarial examples by explicitly characterizing the shared vulnerable regions between SAM and downstream models through a parametric simplicial complex. Our goal is to identify such complexes within adversarially potent regions by iterative vertex-wise refinement.A lightweight domain re-adaptation strategy is introduced to bridge domain divergence using minimal reference data. Notably, VeSCA leverages only the encoder of SAM, which mitigates overfitting issue, and generates consistently transferable adversarial examples by random simplicial complex sampling. Extensive experiments demonstrate that VeSCA achieves performance improved by 12.7% compared to state-of-the-art methods across three downstream model categories across five domain-specific datasets. Our findings further highlight the downstream model risks posed by SAM's vulnerabilities.