CVPR 2025 Papers

Recent multi-teacher distillation methods have unified the encoders of multiple foundation models into a single encoder, achieving competitive performance on core vision tasks like classification, segmentation, and depth estimation. This led us to ask: Could similar success be achieved when the pool of teachers also includes vision models specialized in diverse tasks across both 2D and 3D perception? In this paper, we define and investigate the problem of heterogeneous teacher distillation, or co-distillation, a challenging multi-teacher distillation scenario where teacher models vary significantly in both (a) their design objectives and (b) the data they were trained on. We explore data-sharing strategies and teacher-specific encoding, and introduce DUNE, a single encoder excelling in 2D vision, 3D understanding, and 3D human perception. Our model achieves performance comparable to that of its larger teachers, sometimes even outperforming them, on their respective tasks. Notably, DUNE surpasses MASt3R in Map-free Visual Relocalization with a much smaller encoder.

Vision-Language (VL) alignment across image and text modalities is a challenging task due to the inherent semantic ambiguity of data with multiple possible meanings. Existing methods typically solve it by learning multiple sub-representation spaces to encode each input data as a set of embeddings, and constraining diversity between whole subspaces to capture diverse semantics for accurate VL alignment. Despite their promising outcomes, existing methods suffer two imperfections: 1) actually, specific semantics is mainly expressed by some local dimensions within the subspace. Ignoring this intrinsic property, existing diversity constraints imposed on the whole subspace may impair diverse embedding learning; 2) multiple embeddings are inevitably introduced, sacrificing computational and storage efficiency. In this paper, we propose a simple yet effective Diverse and Hybrid Set-embeddings learning framework (DH-Set), which is distinct from prior work in three aspects. DH-Set 1) devises a novel semantic importance dissecting method to focus on key local dimensions within each subspace; and thereby 2) not only imposes finer-grained diversity constraint to improve the accuracy of diverse embedding learning, 3) but also mixes key dimensions of all subspaces into the single hybrid embedding to boost inference efficiency. Extensive experiments on various benchmarks and model backbones show the superiority of DH-Set over state-of-the-art methods, achieving substantial 2.3%-14.7% rSum improvements while lowering computational and storage complexity.

The automatic generation of radiology reports has emerged as a promising solution to reduce a time-consuming task and accurately capture critical disease-relevant findings in X-ray images. Previous approaches for radiology report generation have shown impressive performance. However, there remains significant potential to improve accuracy by ensuring that retrieved reports contain disease-relevant findings similar to those in the X-ray images and by refining generated reports. In this study, we propose a Disease-aware image-text Alignment and self-correcting Re-alignment for Trustworthy radiology report generation (DART) framework. In the first stage, we generate initial reports based on image-to-text retrieval with disease-matching, embedding both images and texts in a shared embedding space through contrastive learning. This approach ensures the retrieval of reports with similar disease-relevant findings that closely align with the input X-ray images. In the second stage, we further enhance the initial reports by introducing a self-correction module that re-aligns them with the X-ray images. Our proposed framework achieves state-of-the-art results on two widely used benchmarks, surpassing previous approaches in both report generation and clinical efficacy metrics, thereby enhancing the trustworthiness of radiology reports.

Direct Preference Optimization (DPO) has been proposed as an effective and efficient alternative to reinforcement learning from human feedback (RLHF). In this paper, we propose a novel and enhanced version of DPO based on curriculum learning for text-to-image generation. Our method is divided into two training stages. First, a ranking of the examples generated for each prompt is obtained by employing a reward model. Then, increasingly difficult pairs of examples are sampled and provided to a text-to-image generative (diffusion or consistency) model. Generated samples that are far apart in the ranking are considered to form easy pairs, while those that are close in the ranking form hard pairs. In other words, we use the rank difference between samples as a measure of difficulty. The sampled pairs are split into batches according to their difficulty levels, which are gradually used to train the generative model. Our approach, Curriculum DPO, is compared against state-of-the-art fine-tuning approaches on nine benchmarks, outperforming the competing methods in terms of text alignment, aesthetics and human preference. Our code is available at https://github.com/CroitoruAlin/Curriculum-DPO.

Multi-modal 3D object understanding has gained significant attention, yet current approaches often assume complete data availability and rigid alignment across all modalities. We present CrossOver, a novel framework for cross-modal 3D scene understanding via flexible, scene-level modality alignment. Unlike traditional methods that require aligned modality data for every object instance, CrossOver learns a unified, modality-agnostic embedding space for scenes by aligning modalities -- RGB images, point clouds, CAD models, floorplans, and text descriptions -- with relaxed constraints and without explicit object semantics. Leveraging dimensionality-specific encoders, a multi-stage training pipeline, and emergent cross-modal behaviors, CrossOver supports robust scene retrieval and object localization, even with missing modalities. Evaluations on ScanNet and 3RScan datasets show its superior performance across diverse metrics, highlighting CrossOver's adaptability for real-world applications in 3D scene understanding.

Vision-language models (VLMs) have shown remarkable advancements in multimodal reasoning tasks. However, they still often generate inaccurate or irrelevant responses due to issues like hallucinated image understandings or unrefined reasoning paths. To address these challenges, we introduce Critic-V, a novel framework inspired by the Actor-Critic paradigm to boost the reasoning capability of VLMs. This framework decouples the reasoning process and critic process by integrating two independent components: the Reasoner, which generates reasoning paths based on visual and textual inputs, and the Critic, which provides constructive critique to refine these paths. In this approach, the Reasoner generates reasoning responses according to text prompts, which can evolve iteratively as a policy based on feedback from the Critic. This interaction process was theoretically driven by a reinforcement learning framework where the Critic offers natural language critiques instead of scalar rewards, enabling more nuanced feedback to boost the Reasoner's capability on complex reasoning tasks. The Critic model is trained using Direct Preference Optimization (DPO), leveraging a preference dataset of critiques ranked by Rule-based Reward (RBR) to enhance its critic capabilities. Evaluation results show that the Critic-V framework significantly outperforms existing methods, including GPT-4V, on 5 out of 8 benchmarks, especially regarding reasoning accuracy and efficiency. Combining a dynamic text-based policy for the Reasoner and constructive feedback from the preference-optimized Critic enables a more reliable and context-sensitive multimodal reasoning process. Our approach provides a promising solution to enhance the reliability of VLMs, improving their performance in real-world reasoning-heavy multimodal applications such as autonomous driving and embodied intelligence.

Multimodal RLHF usually happens after supervised finetuning (SFT) stage to continually improve vision-language models' (VLMs) comprehension. Conventional wisdom holds its superiority over continual SFT during this preference alignment stage. In this paper, we observe that the inherent value of multimodal RLHF lies in its negative supervision, the logit of the rejected responses. We thus propose a novel negative supervised finetuning (nSFT) approach that fully excavates these information resided. Our nSFT disentangles this negative supervision in RLHF paradigm, and continually aligns VLMs with a simple SFT loss. This is more memory efficient than multimodal RLHF where 2 (e.g., DPO) or 4 (e.g., PPO) large VLMs are strictly required. The effectiveness of nSFT is rigorously proved by comparing it with various multimodal RLHF approaches, across different dataset sources, base VLMs and evaluation metrics. Besides, fruitful of ablations are provided to support our hypothesis. Code will be found in https://github.com/Kevinz-code/nSFT/.

The Transformer architecture has revolutionized various fields since it was proposed, where positional encoding plays an essential role in effectively capturing sequential order and context. Therefore, Rotary Positional Encoding (RoPE) was proposed to alleviate these issues, which integrates positional information by rotating the embeddings in the attention mechanism. However, RoPE utilizes manually defined rotation matrices, a design choice that favors computational efficiency but limits the model's flexibility and adaptability. In this work, we propose ComRoPE, which generalizes RoPE by defining it in terms of trainable commuting angle matrices. Specifically, we demonstrate that pairwise commutativity of these matrices is essential for RoPE to achieve scalability and positional robustness. We formally define the RoPE Equation, which is an essential condition that ensures consistent performance with position offsets. Based on the theoretical analysis, we present two types of trainable commuting angle matrices as sufficient solutions to the RoPE equation, which significantly improve performance, surpassing the current state-of-the-art method by 1.6% at training resolution and 2.9% at higher resolution on the ImageNet-1K dataset. Furthermore, our framework shows versatility in generalizing to existing RoPE formulations and offering new insights for future positional encoding research. To ensure reproducibility, the source code and instructions are available at https://github.com/Longin-Yu/ComRoPE

Despite recent advances in deep texture recognition, existing methods still lack representational diversity and struggle to capture and preserve discriminative cues across stages of representation hierarchies. Moreover, many rely on loss formulations that prioritize recognition accuracy while overlooking spatial coherence and statistical consistency in the feature space. To address these issues, we propose three key innovations: Stochastic Local Texture Masking (SLTM), a regularization strategy that randomly occludes small texture patches to promote the learning of broader spatial and contextual dependencies; the Chebyshev Attention Depth Permutation Texture Network (CAPTN), a novel architecture that learns expressive and persistent Latent Texture Attribute (LTA) representations. CAPTN integrates a Texture Frequency Attention (TFA) module that generates LTAs and enables frequency-aware interpretability, a Dual Depth Permutation (D2P) module to expose complementary channel adjacency patterns, and Learnable Chebyshev Polynomials (LCPs) to model high-order orderless LTA transformations via recursive Chebyshev basis expansion; and a Latent Texture Attribute Loss that jointly optimizes classification accuracy, statistical alignment, and spatial fidelity. CAPTN supports end-to-end training without relying on fine-tuned CNN backbones and achieves state-of-the-art performance on several texture and material recognition benchmarks. (Code: https://github.com/RavishankarEvani/CAPTN)

Traditional text-based person retrieval (TPR) relies on a single-shot text as query to retrieve the target person, assuming that the query completely captures the user's search intent. However, in real-world scenarios, it can be challenging to ensure the information completeness of such single-shot text. To address this limitation, we propose chat-based person retrieval (ChatPR), a new paradigm that takes an interactive dialogue as query to perform the person retrieval, engaging the user in conversational context to progressively refine the query for accurate person retrieval. The primary challenge in ChatPR is the lack of available dialogue-image paired data. To overcome this challenge, we establish ChatPedes, the first dataset designed for ChatPR, which is constructed by leveraging large language models to automate the question generation and simulate user responses. Additionally, to bridge the modality gap between dialogues and images, we propose a dialogue-refined cross-modal alignment (DiaNA) framework, which leverages two adaptive attribute refiners to bottleneck the conversational and visual information for fine-grained cross-modal alignment. Moreover, we propose a dialogue-specific data augmentation strategy, random round retaining, to further enhance the model's generalization ability across varying dialogue lengths. Extensive experiments demonstrate that DiaNA significantly outperforms existing TPR approaches, highlighting the effectiveness of conversational interactions for person retrieval.

Recent video-language alignment models are trained on sets of videos, each with an associated positive caption and a negative caption generated by large language models. A problem with this procedure is that negative captions may introduce linguistic biases, i.e., concepts are seen only as negatives and never associated with a video. While a solution would be to collect videos for the negative captions, existing databases lack the fine-grained variations needed to cover all possible negatives. In this work, we study whether synthetic videos can help to overcome this issue. Our preliminary analysis with multiple generators shows that, while promising on some tasks, synthetic videos harm the performance of the model on others. We hypothesize this issue is linked to noise (semantic and visual) in the generated videos and develop a method, SynViTA, that accounts for those. SynViTA dynamically weights the contribution of each synthetic video based on how similar its target caption is w.r.t. the real counterpart. Moreover, a semantic consistency loss makes the model focus on fine-grained differences across captions, rather than differences in video appearance. Experiments show that, on average, SynViTA improves over existing methods on VideoCon test sets and SSv2-Temporal, SSv2-Events, and ATP-Hard benchmarks, being a first promising step for using synthetic videos when learning video-language models.

In this work, we introduce Camouflage Anything, a novel and robust approach to generate camouflaged datasets. To the best of our knowledge, we are the first to apply Controlled Out-painting and Representation Engineering for generating realistic camouflaged images with an objective to hide any segmented object coming from a generic or salient database. Our proposed method uses a novel control design to out-paint a given segmented object, with a camouflaged background. We also uncover the role of representation engineering in enhancing the quality of generated camouflage datasets. We address the limitations of existing metrics FID and KID in capturing the 'camouflage quality' ,by proposing a novel metric namely, CamOT. CamOT uses Optimal Transport between foreground & background Gaussian Mixture Models (GMM) of concerned camouflaged object to assign an image quality score. Furthermore, we conduct LoRA-based fine-tuning of the robust BiRefNet baseline with our generated camouflaged datasets, leading to notable improvements in camouflaged object segmentation accuracy. The experimental results showcase the efficacy and potential of Camouflage Anything, outperforming existing methods in camouflaged generation tasks.

We present a marker-based geometric estimation framework for the absolute pose of a camera by analyzing the 1D observations in a single radially distorted pixel scanline.We leverage a pair of known co-planar pencils of lines, along with lens distortion parameters, to propose an ensemble of solvers exploring the space of estimation strategies applicable to our setup.First, we present a minimal algebraic solver requiring only six measurements and yielding eight solutions, which relies on the intersection of two conics defined by one of the pencils of lines.Then, we present a unique closed-form geometric solver from seven measurements.Finally, we present an homography-based formulation amenable to linear least-squares from eight or more measurements.Our geometric framework constitutes a theoretical analysis on the minimum geometric context necessary to solve in closed form for the absolute pose of a single camera from a single radially distorted scanline.

Aligning text-to-image (T2I) diffusion models with prefer-ence optimization is valuable for human-annotated datasets, but the heavy cost of manual data collection limits scalability. Using reward models offers an alternative, however, current preference optimization methods fall short in exploiting the rich information, as they only consider pairwise preference distribution. Furthermore, they lack generalization to multi-preference scenarios and struggle to handle inconsistencies between rewards. To address this, we present Calibrated Preference Optimization (CaPO), a novel method to align T2I diffusion models by incorporating the general preference from multiple reward models without human annotated data. The core of our approach involves a reward calibration method to approximate the general preference by computing the expected win-rate against the samples generated by the pretrained models. Additionally, we propose a frontier-based pair selection method that effectively manages the multi-preference distribution by selecting pairs from Pareto frontiers. Finally, we use regression loss to fine-tune diffusion models to match the difference between calibrated rewards of a selected pair. Experimental results show that CaPO consistently outperforms prior methods, such as Direct Preference Optimization (DPO), in both single and multi-reward settings validated by evaluation on T2I benchmarks, including GenEval and T2I-Compbench.

Recent advances in audio-visual learning have shown promising results in learning representations across modalities. However, most approaches rely on global audio representations that fail to capture fine-grained temporal correspondences with visual frames.Additionally, existing methods often struggle with conflicting optimization objectives when trying to jointly learn reconstruction and cross-modal alignment. In this work, we propose CAV-MAE Sync as a simple yet effective extension of the original CAV-MAE framework for self-supervised audio-visual learning. We address three key challenges: First, we tackle the granularity mismatch between modalities by treating audio as a temporal sequence aligned with video frames, rather than using global representations. Second, we resolve conflicting optimization goals by separating contrastive and reconstruction objectives through dedicated global tokens. Third, we improve spatial localization by introducing learnable register tokens that reduce semantic load on patch tokens. We evaluate the proposed approach on AudioSet, VGG Sound, and the ADE20K Sound dataset on zero-shot retrieval, classification and localization tasks demonstrating state-of-the-art performance and outperforming more complex architectures. Code available at https://github.com/edsonroteia/cav-mae-sync.

Creating CAD digital twins from the physical world is crucial for manufacturing, design, and simulation. However, current methods typically rely on costly 3D scanning with labor-intensive post-processing. To provide a user-friendly design process, we explore the problem of reverse engineering from unconstrained real-world CAD images that can be easily captured by users of all experiences. However, the scarcity of real-world CAD data poses challenges in directly training such models. To tackle these challenges, we propose CADCrafter, an image-to-parametric CAD model generation framework that trains solely on synthetic textureless CAD data while testing on real-world images. To bridge the significant representation disparity between images and parametric CAD models, we introduce a geometry encoder to accurately capture diverse geometric features. Moreover, the texture-invariant properties of the geometric features can also facilitate the generalization to real-world scenarios. Since compiling CAD parameter sequences into explicit CAD models is a non-differentiable process, the network training inherently lacks explicit geometric supervision. To impose geometric validity constraints, we employ direct preference optimization (DPO) to fine-tune our model with the automatic code checker feedback on CAD sequence quality. Furthermore, we collected a real-world dataset, comprised of multi-view images and corresponding CAD command sequence pairs, to evaluate our method. Experimental results demonstrate that our approach can robustly handle real unconstrained CAD images, and even generalize to unseen general objects.

High-dynamic scene optical flow is a challenging task, which suffers spatial blur and temporal discontinuous motion due to large displacement in frame imaging, thus deteriorating the spatiotemporal feature of optical flow. Typically, existing methods mainly introduce event camera to directly fuse the spatiotemporal features between the two modalities. However, this direct fusion is ineffective, since there exists a large gap due to the heterogeneous data representation between frame and event modalities. To address this issue, we explore a common-latent space as an intermediate bridge to mitigate the modality gap. In this work, we propose a novel common spatiotemporal fusion between frame and event modalities for high-dynamic scene optical flow, including visual boundary localization and motion correlation fusion. Specifically, in visual boundary localization, we figure out that frame and event share the similar spatiotemporal gradients, whose similarity distribution is consistent with the extracted boundary distribution. This motivates us to design the common spatiotemporal gradient to constrain the reference boundary localization. In motion correlation fusion, we discover that the frame-based motion possesses spatially dense but temporally discontinuous correlation, while the event-based motion has spatially sparse but temporally continuous correlation. This inspires us to use the reference boundary to guide the complementary motion knowledge fusion between the two modalities. Moreover, common spatiotemporal fusion can not only relieve the cross-modal feature discrepancy, but also make the fusion process interpretable for dense and continuous optical flow. Extensive experiments have been performed to verify the superiority of the proposed method.

Human image generation is a key focus in image synthesis due to its broad applications, but even slight inaccuracies in anatomy, pose, or details can compromise realism. To address these challenges, we explore Direct Preference Optimization (DPO), which trains models to generate preferred (winning) images while diverging from non-preferred (losing) ones. However, conventional DPO methods use generated images as winning images, limiting realism. To overcome this limitation, we propose an enhanced DPO approach that incorporates high-quality real images as winning images, encouraging outputs to resemble real images rather than generated ones. However, implementing this concept is not a trivial task. Therefore, our approach, HG-DPO (Human image Generation through DPO), employs a novel curriculum learning framework that gradually improves the output of the model toward greater realism, making training more feasible. Furthermore, HG-DPO effectively adapts to personalized text-to-image tasks, generating high-quality and identity-specific images, which highlights the practical value of our approach.