CVPR 2025 Papers

High-quality semantic segmentation relies on three key capabilities: global context modeling, local detail encoding, and multi-scale feature extraction. However, recent methods struggle to possess all these capabilities simultaneously. Hence, we aim to empower segmentation networks to simultaneously carry out efficient global context modeling, high-quality local detail encoding, and rich multi-scale feature representation for varying input resolutions. In this paper, we introduce SegMAN, a novel linear-time model comprising a hybrid feature encoder dubbed SegMAN Encoder, and a decoder based on state space models. Specifically, the SegMAN Encoder synergistically integrates sliding local attention with dynamic state space models, enabling highly efficient global context modeling while preserving fine-grained local details. Meanwhile, the MMSCopE module in our decoder enhances multi-scale context feature extraction and adaptively scales with the input resolution. Our SegMAN-B Encoder achieves 85.1% ImageNet-1k accuracy (+1.5% over VMamba-S with fewer parameters). When paired with our decoder, the full SegMAN-B model achieves 52.6% mIoU on ADE20K (+1.6% over SegNeXt-L with 15% fewer GFLOPs), 83.8% mIoU on Cityscapes (+2.1% over SegFormer-B3 with half the GFLOPs), and 1.6% higher mIoU than VWFormer-B3 on COCO-Stuff with lower GFLOPs. Our code is available at https://github.com/yunxiangfu2001/SegMAN.

While MLLMs have demonstrated adequate image understanding capabilities, they still struggle with pixel-level comprehension, limiting their practical applications. Current evaluation tasks like VQA and visual grounding remain too coarse to assess fine-grained pixel comprehension accurately. Though segmentation is foundational for pixel-level understanding, existing methods often require MLLMs to generate implicit tokens, decoded through external pixel decoders. This approach disrupts the MLLM's text output space, potentially compromising language capabilities and reducing flexibility and extensibility, while failing to reflect the model's intrinsic pixel-level understanding.Thus, We introduce the Human-Like Mask Annotation Task (HLMAT), a new paradigm where MLLMs mimic human annotators using interactive segmentation tools. Modeling segmentation as a multi-step Markov Decision Process, HLMAT enables MLLMs to iteratively generate text-based click points, achieving high-quality masks without architectural changes or implicit tokens. Through this setup, we develop SegAgent, a model fine-tuned on human-like annotation trajectories, which achieves performance comparable to SOTA methods and supports additional tasks like mask refinement and annotation filtering.HLMAT provides a protocol for assessing fine-grained pixel understanding in MLLMs and introduces a vision-centric, multi-step decision-making task that facilitates exploration of MLLMs' visual reasoning abilities. Our adaptations of policy improvement method StaR and PRM guided tree search further enhance model robustness in complex segmentation tasks, laying a foundation for future advancements in fine-grained visual perception and multi-step decision-making for MLLMs.

Domain generalization aims to learn a model from multiple training domains and generalize it to unseen test domains. Recent theory has shown that seeking the deep models, whose parameters lie in the flat minima of the loss landscape, can significantly reduce the out-of-domain generalization error. However, existing methods often neglect the consistency of loss landscapes in different domains, resulting in models that are not simultaneously in the optimal flat minima in all domains, which limits their generalization ability. To address this issue, this paper proposes an iterative Self-Feedback Training (SFT) framework to seek consistent flat minima that are shared across different domains by progressively refining loss landscapes during training. It alternatively generates a feedback signal by measuring the inconsistency of loss landscapes in different domains and refines these loss landscapes for greater consistency using this feedback signal. Benefiting from the consistency of the flat minima within these refined loss landscapes, our SFT helps achieve better out-of-domain generalization. Extensive experiments on DomainBed demonstrate superior performances of SFT when compared to state-of-the-art sharpness-aware methods and other prevalent DG baselines. On average across five DG benchmarks, SFT surpasses the sharpness-aware minimization by 2.6% with ResNet-50 and 1.5% with ViT-B/16, respectively.

Multimodal sentiment analysis has attracted extensive research attention as increasing users share images and texts to express their emotions and opinions on social media. Collecting large amounts of labeled sentiment data is an expensive and challenging task due to the high cost of labeling and unavoidable label ambiguity. Semi-supervised learning (SSL) is explored to utilize the extensive unlabeled data to alleviate the demand for annotation. However, different from typical multimodal tasks, the inconsistent sentiment between image and text leads to the sub-optimal performance of SSL algorithms. To address this issue, we propose SCDR, the first semi-supervised image-text sentiment recognition framework. To better utilize the discriminative features of each modality, we decouple features into common and private parts and then use the private features to train unimodal classifiers for enhanced modality-specific sentiment representation. Considering the complex relation between modalities, we devise a modal selection-based attention module that adaptively assesses the dominant sentiment modality at the sample level to guide the fusion of multimodal representations. Furthermore, to prevent the model predictions from overly relying on common features under the guidance of multimodal labels, we design a pseudo-label filtering strategy based on the matching degree of prediction and dominant modality. Extensive experiments and comparisons on five publicly available datasets demonstrate that SCDR outperforms state-of-the-art methods. The code is provided in the supplementary material and will be released to the public.

Natural language goes beyond dryly describing visual content. It contains rich abstract concepts to express feeling, creativity and properties that cannot be directly perceived. Yet, current research in Vision Language Models (VLMs) has not shed light on abstract-oriented language.Our research breaks new ground by uncovering its wide presence and under-estimated value, with extensive analysis.Particularly, we focus our investigation on the fashion domain, a highly-representative field with abstract expressions. By analyzing recent large-scale multimodal fashion datasets, we find that abstract terms have a dominant presence, rivaling the concrete ones, providing novel information, and being useful in the retrieval task. However, a critical challenge emerges: current general-purpose or fashion-specific VLMs are pre-trained with databases that lack sufficient abstract words in their text corpora, thus hindering their ability to effectively represent abstract-oriented language. We propose a training-free and model-agnostic method, Abstract-to-Concrete Translator (ACT), to shift abstract representations towards well-represented concrete ones in the VLM latent space, using pre-trained models and existing multimodal databases.On the text-to-image retrieval task, despite being training-free, ACT outperforms the fine-tuned VLMs in both same- and cross-dataset settings, exhibiting its effectiveness with a strong generalization capability. Moreover, the improvement introduced by ACT is consistent with various VLMs, making it a plug-and-play solution.

Deep learning models for 3D data have shown to be vulnerable to adversarial attacks, which have received increasing attention in various safety-critical applications such as autonomous driving and robotic navigation. Existing 3D attackers mainly put effort into attacking the simple 3D classification model by perturbing point cloud objects in the white/black-box setting. However, real-world 3D applications focus on tackling more complicated scene-based data while sharing no information about the model parameters and logits with users. Therefore, directly applying previous naive 3D attack methods to these applications does not work. To this end, this paper attempts to address the challenging hard-label 3D scene attack with access only to the input/output of the 3D models. To make the attack effective and stealthy, we propose to generate universal adversarial objects, which will mislead scene-aware 3D models to predict attacker-chosen labels whenever these objects are placed on any scene input. Specifically, we inject an imperceptible object trigger with further perturbations into all scenes and learn to mislead their reasoning by only querying the 3D model. We start by initializing the trigger pattern with a realistic object and searching for an appropriate location to place it naturally in the scene data. Then, we design a novel weighted gradient estimation strategy to perturb the object trigger with additive slight noise to make them adversarial in an iterative optimization procedure. Extensive experiments demonstrate that our attack can achieve superior performance on seven 3D models and three scene-based datasets, with satisfactory adversarial imperceptibility and strong resistance to defense methods.

The CLIP model has demonstrated significant advancements in aligning visual and language modalities through large-scale pre-training on image-text pairs, enabling strong zero-shot classification and retrieval capabilities on various domains. However, CLIP's training remains computationally intensive, with high demands on both data processing and memory. To address these challenges, recent masking strategies have emerged, focusing on the selective removal of image patches to improve training efficiency. Although effective, these methods often compromise key semantic information, resulting in suboptimal alignment between visual features and text descriptions.In this work, we present a concise yet effective approach called Patch Generation-to-Selection (CLIP-PGS) to enhance CLIP's training efficiency while preserving critical semantic content. Our method introduces a gradual masking process in which a small set of candidate patches is first pre-selected as potential mask regions. Then, we apply Sobel edge detection across the entire image to generate an edge mask that prioritizes the retention of the primary object areas. Finally, similarity scores between the candidate mask patches and their neighboring patches are computed, with optimal transport normalization refining the selection process to ensure a balanced similarity matrix.Our approach, CLIP-PGS, sets new state-of-the-art results in zero-shot classification and retrieval tasks, achieving superior performance in robustness evaluation and language compositionality benchmarks.

We present a unified model capable of simultaneously grounding both spoken language and non-speech sounds within a visual scene, addressing key limitations in current audio-visual grounding models. Existing approaches are typically limited to handling either speech or non-speech sounds independently, or at best, together but sequentially without mixing. This limitation prevents them from capturing the complexity of real-world audio sources that are often mixed. Our approach introduces a "mix-and-separate" framework with audio-visual alignment objectives that jointly learn correspondence and disentanglement using mixed audio. Through these objectives, our model learns to produce distinct embeddings for each audio type, enabling effective disentanglement and grounding across mixed audio sources.Additionally, we created a new dataset to evaluate simultaneous grounding of mixed audio sources, demonstrating that our model outperforms prior methods. Our approach also achieves state-of-the-art performance in standard segmentation and cross-modal retrieval tasks, highlighting the benefits of our mix-and-separate approach.

Large multimodal models (LMMs) typically process visual inputs with uniform resolution across the entire field of view, leading to inefficiencies when non-critical image regions are processed as precisely as key areas. Inspired by the human visual system's foveated approach, we apply a sampling method to leading architectures such as MDETR, BLIP2, InstructBLIP, LLaVA, and ViLT, and evaluate their performance with variable (foveated) resolution inputs. Results show that foveated sampling boosts accuracy in visual tasks like question answering and object detection under tight pixel budgets, improving performance by up to 2.7% on the GQA dataset, 2.1% on SEED-Bench, and 2.0% on VQAv2 compared to uniform sampling. Furthermore, we show that indiscriminate resolution increases yield diminishing returns, with models achieving up to 80% of their full capability using just 3% of the pixels, even on complex tasks. Foveated sampling prompts more human-like processing within models, such as neuronal selectivity and globally acting self-attention in vision transformers. This paper provides a foundational analysis of foveated sampling's impact on existing models, suggesting that more efficient architectural adaptations, mimicking human visual processing, are a promising research venue for the community. Potential applications of our findings center low power minimal bandwidth devices (such as UAVs and edge devices), where compact and efficient vision is critical.

Recent advancements in multimodal large language models (MLLMs) have significantly improved performance in visual question answering. However, they often suffer from hallucinations. In this work, hallucinations are categorized into two main types: initial hallucinations and snowball hallucinations. We argue that adequate contextual information can be extracted directly from the token interaction process. Inspired by causal inference in decoding strategy, we propose to leverage causal masks to establish information propagation between multimodal tokens. The hypothesis is that insufficient interaction between those tokens may lead the model to rely on outlier tokens, overlooking dense and rich contextual cues. Therefore, we propose to intervene in the propagation process by tackling outlier tokens to enhance in-context inference. With this goal, we present FarSight, a versatile plug-and-play decoding strategy to reduce attention interference from outlier tokens merely by optimizing the causal mask. The heart of our method is effective token propagation. We design an attention register structure within the upper triangular matrix of the causal mask, dynamically allocating attention capture attention diverted to outlier tokens. Moreover, a positional awareness encoding method with a diminishing masking rate is proposed, allowing the model to attend to further preceding tokens, especially for video sequence tasks. With extensive experiments, FarSight demonstrates significant hallucination-mitigating performance across different MLLMs on both image and video benchmarks, proving its effectiveness.

We present a snapshot imaging technique for recovering 3D surrounding views of miniature scenes. Due to their intricacy, miniature scenes with objects sized in millimeters are difficult to reconstruct, yet miniatures are common in life and their 3D digitalization is desirable. We design a catadioptric imaging system with a single camera and eight pairs of planar mirrors for snapshot 3D reconstruction from a dollhouse perspective. We place paired mirrors on nested pyramid surfaces for capturing surrounding multi-view images in a single shot. Our mirror design is customizable based on the size of the scene for optimized view coverage. We use the 3D Gaussian Splatting (3DGS) representation for scene reconstruction and novel view synthesis. We overcome the challenge posed by our sparse view input by integrating visual hull-derived depth constraint. Our method demonstrates state-of-the-art performance on a variety of synthetic and real miniature scenes.

Video restoration poses non-trivial challenges in maintaining fidelity while recovering temporally consistent details from unknown degradations in the wild. Despite recent advances in diffusion-based restoration, these methods often face limitations in generation capability and sampling efficiency. In this work, we present \text SeedVR , a diffusion transformer designed to handle real-world video restoration with arbitrary length and resolution. The core design of SeedVR lies in the shifted window attention that facilitates effective restoration on long video sequences. SeedVR further supports variable-sized windows near the boundary of both spatial and temporal dimensions, overcoming the resolution constraints of traditional window attention. Equipped with contemporary practices, including causal video autoencoder, mixed image and video training, and progressive training, SeedVR achieves highly-competitive performance on both synthetic and real-world benchmarks, as well as AI-generated videos. Extensive experiments demonstrate SeedVR's superiority over existing methods for generic video restoration.

3D Visual Grounding (3DVG) aims to locate objects in 3D scenes based on textual descriptions, essential for applications like augmented reality and robotics. Traditional 3DVG approaches rely on annotated 3D datasets and predefined object categories, limiting scalability and adaptability. To overcome these limitations, we introduce SeeGround, a zero-shot 3DVG framework leveraging 2D Vision-Language Models (VLMs) trained on large-scale 2D data. SeeGround represents 3D scenes as a hybrid of query-aligned rendered images and spatially enriched text descriptions, bridging the gap between 3D data and 2D-VLMs input formats. We propose two modules: the Perspective Adaptation Module, which dynamically selects viewpoints for query-relevant image rendering, and the Fusion Alignment Module, which integrates 2D images with 3D spatial descriptions to enhance object localization. Extensive experiments on ScanRefer and Nr3D demonstrate that our approach outperforms existing zero-shot methods by large margins. Notably, we exceed weakly supervised methods and rival some fully supervised ones, outperforming previous SOTA by 7.7% on ScanRefer and 7.1% on Nr3D, showcasing its effectiveness in complex 3DVG task. Project website (with demo and code): https://seeground.github.io/.

Significant advances have been made in the sampling efficiency of diffusion and flow matching models, driven by Consistency Distillation (CD), which trains a student model to mimic the output of a teacher model at a later timestep. However, we found that the knowledge discrepancy between student and teacher varies significantly across different timesteps, leading to suboptimal performance in CD.To address this issue, we propose the Curriculum Consistency Model (CCM), which stabilizes and balances the knowledge discrepancy across timesteps. Specifically, we regard the distillation process at each timestep as a curriculum and introduce a metric based on the Peak Signal-to-Noise Ratio (PSNR) to quantify the knowledge discrepancy of this curriculum, then ensure that the curriculum maintains consistent knowledge discrepancy across different timesteps by having the teacher model iterate more steps when the noise intensity is low.Our method achieves competitive single-step sampling Frechet Inception Distance (FID) scores of 1.64 on CIFAR-10 and 2.18 on ImageNet 64x64.Moreover, we have extended our method to large-scale text-to-image models and confirmed that it generalizes well to both diffusion models (Stable Diffusion XL) and flow matching models (Stable Diffusion 3). The generated samples demonstrate improved image-text alignment and semantic structure since CCM enlarges the distillation step at large timesteps and reduces the accumulated error.

The advancement of Generative Adversarial Networks (GANs) and diffusion models significantly enhances the realism of synthetic images, driving progress in image processing and creative design. However, this progress also necessitates the development of effective detection methods, as synthetic images become increasingly difficult to distinguish from real ones. This difficulty leads to societal issues, such as the spread of misinformation, identity theft, and online fraud. While previous detection methods perform well on public benchmarks, they struggle with our benchmark, FakeART, particularly when dealing with the latest models and cross-domain tasks (e.g., photo-to-painting). To address this challenge, we develop a new synthetic image detection technique based on color distribution. Unlike real images, synthetic images often exhibit uneven color distribution. By employing color quantization and restoration techniques, we analyze the color differences before and after image restoration. We discover and prove that these differences closely relate to the uniformity of color distribution. Based on this finding, we extract effective color features and combine them with image features to create a detection model with only 1.4 million parameters. This model achieves state-of-the-art results across various evaluation benchmarks, including the challenging FakeART dataset.

In this paper, we introduce SearchDet, a training-free long-tail object detection framework that significantly enhances open-vocabulary object detection performance. SearchDet retrieves a set of positive and negative images of an object to ground, embeds these images, and computes an input image--weighted query which is used to detect the desired concept in the image. Our proposed method is simple and training-free, yet achieves over 16.81% mAP improvement on ODinW and 59.85% mAP improvement on LVIS compared to state-of-the-art models such as GroundingDINO. We further show that our approach of basing object detection on a set of Web-retrieved exemplars is stable with respect to variations in the exemplars, suggesting a path towards eliminating costly data annotation and training procedures.

Denoising diffusion probabilistic models have achieved significant success in point cloud generation, enabling numerous downstream applications, such as generative data augmentation and 3D model editing. However, little attention has been given to generating point clouds with point-wise segmentation labels, as well as to developing evaluation metrics for this task. Therefore, in this paper, we present SeaLion, a novel diffusion model designed to generate high-quality and diverse point cloud with fine-grained segmentation labels. Specifically, we introduce the semantic part-aware latent point diffusion technique, which leverages the intermediate features of the generative models to jointly predict the noise for perturbed latent points and associated part segmentation labels during the denoising process, and subsequently decodes the latent points to point clouds conditioned on part segmentation labels. To effectively evaluate the quality of generated point clouds, we introduce a novel point cloud pairwise distance calculation method named part-aware Chamfer distance (p-CD). This method enables existing metrics, such as 1-NNA, to measure both the local structural quality and inter-part coherence of generated point clouds. Experiments on the large-scale synthetic dataset ShapeNet and real-world medical dataset IntrA, demonstrate that SeaLion achieves remarkable performance in generation quality and diversity, outperforming the existing state-of-the-art model, DiffFacto, by 13.33% and 6.52% on 1-NNA (p-CD) across the two datasets. Experimental analysis shows that SeaLion can be trained semi-supervised, thereby reducing the demand for labeling efforts. Lastly, we validate the applicability of SeaLion in generative data augmentation for training segmentation models and the capability of SeaLion to serve as a tool for part-aware 3D shape editing.

Underwater (UW) robotics applications require depth and restored images simultaneously in real-time, irrespective of whether the UW images are captured in good lighting conditions or not. Most of the UW image restoration and depth estimation methods have been devised for images under normal lighting. Consequently, they struggle to perform on poorly lit images. Even though artificial illumination can be used when there is insufficient ambient light, it can introduce non-uniform lighting artifacts in the restored images. Hence, the recovery of depth and restored images directly from Low-Light UW (LLUW) images is a critical requirement in marine applications. While a few works have attempted LLUW image restoration, there are no reported works on joint recovery of depth and clean image from LLUW images. We propose a Self-supervised Low-light Underwater Image and Depth recovery network (SelfLUID-Net) for joint estimation of depth and restored image in real-time from a single LLUW image. We have collected an Underwater Low-light Stereo Video (ULVStereo) dataset which is the first-ever UW dataset with stereo pairs of low-light and normally-lit UW images. For the dual tasks of image and depth recovery from a LLUW image, we effectively utilize the stereo data from ULVStereo that provides cues for both depth and illumination-independent clean image. We harness a combination of the UW image formation process, the Retinex model, and constraints enforced by the scene geometry for our self-supervised training. To handle occlusions, we additionally utilize monocular frames from our video dataset and propose a masking scheme to prevent dynamic transients, that do not respect the underlying scene geometry, from misguiding the learning process. Evaluations on five LLUW datasets demonstrate the superiority and generalization ability of our proposed SelfLUID-Net over existing state-of-the-art methods. The dataset ULVStereo is available at https://github.com/nishavarghese15/ULVStereo.