CVPR 2025 Papers

Identifying affordance regions on 3D objects from semantic cues is essential for robotics and human-machine interaction. However, existing 3D affordance learning methods struggle with generalization and robustness due to limited annotated data and a reliance on 3D backbones focused on geometric encoding, which often lack resilience to real-world noise and data corruption. We propose GEAL, a novel framework designed to enhance the generalization and robustness of 3D affordance learning by leveraging large-scale pre-trained 2D models. We employ a dual-branch architecture with Gaussian splatting to establish consistent mappings between 3D point clouds and 2D representations, enabling realistic 2D renderings from sparse point clouds. A granularity-adaptive fusion module and a 2D-3D consistency alignment module further strengthen cross-modal alignment and knowledge transfer, allowing the 3D branch to benefit from the rich semantics and generalization capacity of 2D models. To holistically assess the robustness, we introduce two new corruption-based benchmarks: PIAD-C and LASO-C. Extensive experiments on public datasets and our benchmarks show that GEAL consistently outperforms existing methods across seen and novel object categories, as well as corrupted data, demonstrating robust and adaptable affordance predictions. The code and datasets are publicly available.

We examine the capability of Multimodal Large Language Models (MLLMs) to tackle diverse domains that extend beyond the traditional language and vision tasks these models are typically trained on. Specifically, our focus lies in areas such as Embodied AI, Games, UI Control, and Planning. To this end, we introduce a process of adapting an MLLM to a Generalist Embodied Agent (GEA). GEA is a single unified model capable of grounding itself across these varied domains through a multi-embodiment action tokenizer. GEA is trained with supervised learning on a large dataset of embodied experiences and with online RL in interactive simulators. We explore the data and algorithmic choices necessary to develop such a model. Our findings reveal the importance of training with cross-domain data and online RL for building generalist agents. The final GEA model achieves strong generalization performance to unseen tasks across diverse benchmarks compared to other generalist models and benchmark-specific approaches.

Tremendous progress has been made in deep stereo matching to excel on benchmark datasets through per-domain fine-tuning. However, achieving strong zero-shot generalization - a hallmark of foundation models in other computer vision tasks - remains challenging for stereo matching. We introduce FoundationStereo, a foundation model for stereo depth estimation designed to achieve strong zero shot generalization. To this end, we first construct a large scale (1M stereo pairs) synthetic training dataset featuring large diversity and high photorealism, followed by an automatic self-curation pipeline to remove ambiguous samples. We then design a number of network architecture components to enhance scalability, including a side-tuning feature backbone that adapts rich monocular priors from vision foundation models to mitigate the sim-to-real gap, and long-range context reasoning for effective cost volume filtering. Together, these components lead to strong robustness and accuracy across domains, establishing a new standard in zero-shot stereo depth estimation. Project page: https://nvlabs.github.io/FoundationStereo

Ensuring auditability and verifiability in FL is both challenging and essential to guarantee that local data remains untampered and client updates are trustworthy. Recent FL frameworks assess client contributions through a trusted central server using various client selection and aggregation techniques. However, reliance on a central server can create a single point of failure, making it vulnerable to privacy-centric attacks and limiting its ability to audit and verify client-side data contributions due to restricted access. In addition, data quality and fairness evaluations are often inadequate, failing to distinguish between high-impact contributions and those from low-quality or poisoned data. To address these challenges, we propose Federated Auditable and Verifiable Data valuation (FAVD), a privacy-preserving method that ensures auditability and verifiability of client contributions through data valuation, independent of any central authority or predefined training algorithm. FAVD utilizes shared local data density functions to construct a global density function, aligning data contributions and facilitating effective valuation prior to local model training. This proactive approach improves transparency in data valuation and ensures that only benign updates are generated, even in the presence of malicious data. Further, to mitigate privacy risks associated with sharing data density functions, we add Gaussian noise to each client's local density function before sharing it with the server. We theoretically demonstrate the convergence, auditability, and verifiability of FAVD, along with its resilience against data poisoning threats. Our experiments on five diverse benchmarks, including three medical datasets, show that FAVD achieves significant performance gains, accurate data valuation, and fair client contributions under threat, highlighting its reliability as a trustworthy FL approach.

Learning accurate scene reconstruction without pose priors in neural radiance fields is challenging due to inherent geometric ambiguity. Recent development either relies on correspondence priors for regularization or uses off-the-shelf flow estimators to derive analytical poses. However, the potential for jointly learning scene geometry, camera poses, and dense flow within a unified neural representation remains largely unexplored. In this paper, we present Flow-NeRF, a unified framework that simultaneously optimizes scene geometry, camera poses, and dense optical flow all on-the-fly. To enable the learning of dense flow within the neural radiance field, we design and build a bijective mapping for flow estimation, conditioned on pose. To make the scene reconstruction benefit from the flow estimation, we develop an effective feature enhancement mechanism to pass canonical space features to world space representations, significantly enhancing scene geometry. We validate our model across four important tasks, i.e., novel view synthesis, depth estimation, camera pose prediction, and dense optical flow estimation, using several datasets. Our approach surpasses previous methods in almost all metrics for novel-view view synthesis and depth estimation and yields both qualitatively sound and quantitatively accurate novel-view flow. Our project page is https://zhengxunzhi.github.io/flownerf/.

We present Florence-VL, a new family of multimodal large language models (MLLMs) with enriched visual representations produced by Florence-2, a generative vision foundation model. Unlike the widely used CLIP-style vision transformer trained by contrastive learning, Florence-2 can capture different levels and aspects of visual features, which are more versatile to be adapted to diverse downstream tasks. We propose a novel feature-fusion architecture and an innovative training recipe that effectively integrates Florence-2's visual features into pretrained LLMs, such as Phi 3.5 and LLama 3. In particular, we propose "depth-breath fusion (DBFusion)" to fuse the visual features extracted from different depths and under multiple prompts. Our model training is composed of end-to-end pretraining of the whole model followed by finetuning of the projection layer and the LLM, on a carefully designed recipe of diverse open-source datasets that include high-quality image captions and instruction-tuning pairs. Our quantitative analysis and visualization of Florence-VL's visual features show its advantages over popular vision encoders on vision-language alignment, where the enriched depth and breath play important roles. Florence-VL achieves significant improvements over existing state-of-the-art MLLMs across various multi-modal and vision-centric benchmarks covering general VQA, perception, hallucination, OCR, Chart, knowledge-intensive understanding, etc. To facilitate future research, our models and the complete training recipe are open-sourced.

Recent advances in 3D Gaussian Splatting (3DGS) have demonstrated significant potential over traditional rendering techniques, attracting widespread attention from both industry and academia. However, real-time rendering with 3DGS remains a challenging problem, particularly in large-scale, high-resolution scenes due to the presence of numerous anisotropic Gaussian representations, and it has not been extensively explored. To address this challenge, we introduce FlashGS, an open-source CUDA library with Python bindings, featuring comprehensive algorithm design and optimizations, including redundancy elimination, adaptive scheduling, and efficient pipelining. First, we eliminate substantial redundant computations through precise Gaussian intersection tests, leveraging the intrinsic mechanism of the 3DGS rasterizer. During task partitioning, we propose an adaptive scheduling strategy that accounts for variations in Gaussian size and shape. Additionally, we design a multi-stage pipelining strategy for color computation in the rendering process, further accelerating performance. We conduct an extensive evaluation of FlashGS across a diverse range of synthetic and real-world 3D scenes, encompassing scene sizes of up to 2.7 km^2 cityscape and resolutions of up to over 4K. Our approach improves 3DGS rendering performance by an order of magnitude, achieving an average speedup of 7.2x, and rendering at a minimum of 125.9 FPS, setting a new state-of-the-art in real-time 3DGS rendering. https://github.com/InternLandMark/FlashGS.

Although multi-instance learning (MIL) has succeeded in pathological image classification, it faces the challenge of high inference costs due to processing numerous patches from gigapixel whole slide images (WSIs).To address this, we propose HDMIL, a hierarchical distillation multi-instance learning framework that achieves fast and accurate classification by eliminating irrelevant patches.HDMIL consists of two key components: the dynamic multi-instance network (DMIN) and the lightweight instance pre-screening network (LIPN). DMIN operates on high-resolution WSIs, while LIPN operates on the corresponding low-resolution counterparts.During training, DMIN are trained for WSI classification while generating attention-score-based masks that indicate irrelevant patches.These masks then guide the training of LIPN to predict the relevance of each low-resolution patch.During testing, LIPN first determines the useful regions within low-resolution WSIs, which indirectly enables us to eliminate irrelevant regions in high-resolution WSIs, thereby reducing inference time without causing performance degradation.In addition, we further design the first Chebyshev-polynomials-based Kolmogorov-Arnold classifier in computational pathology, which enhances the performance of HDMIL through learnable activation layers.Extensive experiments on three public datasets demonstrate that HDMIL outperforms previous state-of-the-art methods, e.g., achieving improvements of 3.13% in AUC while reducing inference time by 28.6% on the Camelyon16 dataset.The project will be available.

Multimodal large language models (MLLMs) have demonstrated remarkable capabilities in various tasks. However, effectively evaluating these MLLMs on face perception remains largely unexplored. To address this gap, we introduce FaceBench, a dataset featuring hierarchical multi-view and multi-level attributes specifically designed to assess the comprehensive face perception abilities of MLLMs. Initially, we construct a hierarchical facial attribute structure, which encompasses five views with up to three levels of attributes, totaling over 210 attributes and 700 attribute values. Based on the structure, the proposed FaceBench consists of 49,919 visual question-answering (VQA) pairs for evaluation and 23,841 pairs for fine-tuning. Moreover, we further develop a robust face perception MLLM baseline, Face-LLaVA, by training with our proposed face VQA data. Extensive experiments on various mainstream MLLMs and Face-LLaVA are conducted to test their face perception ability, with results also compared against human performance. The results reveal that, the existing MLLMs are far from satisfactory in understanding the fine-grained facial attributes, while our Face-LLaVA significantly outperforms existing open-source models with a small amount of training data and is comparable to commercial ones like GPT-4o and Gemini. The dataset will be released at https://github.com/CVI-SZU/FaceBench.

Progress in machine understanding of sign languages has been slow and hampered by limited data. In this paper, we present FSboard, an American Sign Language fingerspelling dataset situated in a mobile text entry use case, collected from 147 paid and consenting Deaf signers using Pixel 4A selfie cameras in a variety of environments. Fingerspelling recognition is an incomplete solution that is only one small part of sign language translation, but it could provide some immediate benefit to Deaf/Hard of Hearing signers while more broadly capable technology develops. At >3 million characters in length and >250 hours in duration, FSboard is the largest fingerspelling recognition dataset to date by a factor of >10x. As a simple baseline, we finetune 30 Hz MediaPipe Holistic landmark inputs into ByT5-Small and achieve 11.1% Character Error Rate (CER) on a test set with unique phrases and signers. This quality degrades gracefully when decreasing frame rate and excluding face/body landmarks---plausible optimizations to help with on-device performance---but falls short of human performance measured at 2.2% CER.

Endowing Large Multimodal Models (LMMs) with visual grounding capability can significantly enhance AIs' understanding of the visual world and their interaction with humans. However, existing methods typically fine-tune the parameters of LMMs to learn additional segmentation tokens and overfit grounding and segmentation datasets. Such a design would inevitably cause a catastrophic diminution in the indispensable conversational capability of general AI assistants. In this paper, we comprehensively evaluate state-of-the-art grounding LMMs across a suite of multimodal question-answering benchmarks, observing drastic performance drops that indicate vanishing general knowledge comprehension and weakened instruction following ability. To address this issue, we present F-LMM---grounding frozen off-the-shelf LMMs in human-AI conversations---a straightforward yet effective design based on the fact that word-pixel correspondences conducive to visual grounding inherently exist in the attention mechanism of well-trained LMMs. Using only a few trainable CNN layers, we can translate word-pixel attention weights to mask logits, which a SAM-based mask refiner can further optimise. Our F-LMM neither learns special segmentation tokens nor utilises high-quality grounded instruction-tuning data, but achieves competitive performance on referring expression segmentation and panoptic narrative grounding benchmarks while completely preserving LMMs' original conversational ability. Additionally, with instruction-following ability preserved and grounding ability obtained, F-LMM can be directly applied to complex tasks like reasoning segmentation, grounded conversation generation and visual chain-of-thought reasoning. Our code can be found at https://github.com/wusize/F-LMM.

We present a technique and benchmark dataset for estimating the relative 3D orientation between a pair of Internet images captured in an extreme setting, where the images have limited or non-overlapping field of views. Prior work targeting extreme rotation estimation assume constrained 3D environments and emulate perspective images by cropping regions from panoramic views. However, real images captured in the wild are highly diverse, exhibiting variation in both appearance and camera intrinsics. In this work, we propose a Transformer-based method for estimating relative rotations in extreme real-world settings, and contribute the ExtremeLandmarkPairs dataset, assembled from scene-level Internet photo collections. Our evaluation demonstrates that our approach succeeds in estimating the relative rotations in a wide variety of extreme-view Internet image pairs, outperforming various baselines, including dedicated rotation estimation techniques and contemporary 3D reconstruction methods.

Detecting 3D objects in point clouds plays a crucial role in autonomous driving systems. Recently, advanced multi-modal methods incorporating camera information have achieved notable performance. For a safe and effective autonomous driving system, algorithms that excel not only in accuracy but also in speed and low latency are essential. However, existing algorithms fail to meet these requirements due to the latency and bandwidth limitations of fixed frame rate sensors, e.g., LiDAR and camera. To address this limitation, we introduce asynchronous event cameras into 3D object detection for the first time. We leverage their high temporal resolution and low bandwidth to enable high-speed 3D object detection. Our method enables detection even during inter-frame intervals when synchronized data is unavailable, by retrieving previous 3D information through the event camera. Furthermore, we introduce the first event-based 3D object detection datasets, Ev-Waymo and DSEC-3DOD, both of which include ground truth 3D bounding boxes at 100 FPS, establishing the first benchmarks for event-based 3D detectors. The code and dataset are available at https://github.com/mickeykang16/Ev3DOD.

High-quality open-source datasets are essential for advancing deep neural networks. However, the unauthorized commercial use of these datasets has raised significant concerns about copyright protection. One promising approach is backdoor watermark-based dataset ownership verification (BW-DOV), in which dataset protectors implant specific backdoors into illicit models through dataset watermarking, enabling the tracing of these models through abnormal prediction behaviors. Unfortunately, the targeted nature of these BW-DOV methods can be maliciously exploited, potentially leading to harmful side effects. While existing harmless methods attempt to mitigate these risks, watermarked datasets can still negatively affect prediction results, partially compromising dataset functionality. In this paper, we propose a more harmless backdoor watermark, called EntropyMark, which improves prediction confidence without altering the final prediction results. For this purpose, an entropy-based constraint is introduced to regulate the probability distribution. Specifically, we design an iterative clean-label dataset watermarking framework. Our framework employs gradient matching and adaptive data selection to optimize backdoor injection. In parallel, we introduce a hypothesis test method grounded in entropy inconsistency to verify dataset ownership. Extensive experiments on benchmark datasets demonstrate the effectiveness, transferability, and defense resistance of our approach.

Online continual learning (OCL) seeks to learn new tasks from data streams that appear only once, while retaining knowledge of previously learned tasks. Most existing methods rely on replay, focusing on enhancing memory retention through regularization or distillation. However, they often overlook the adaptability of the model, limiting the ability to learn generalizable and discriminative features incrementally from online training data.To address this, we introduce a plug-and-play module, S6MOD, which can be integrated into most existing methods and directly improve adaptability. Specifically, S6MOD introduces an extra branch after the backbone, where a mixture of discretization selectively adjusts parameters in a selective state space model, enriching selective scan patterns such that the model can adaptively select the most sensitive discretization method for current dynamics.We further design a class-conditional routing algorithm for dynamic, uncertainty-based adjustment and implement a contrastive discretization loss to optimize it. Extensive experiments combining our module with various models demonstrate that S6MOD significantly enhances model adaptability, leading to substantial performance gains and achieving the state-of-the-art results. The code is available at https://github.com/MyToumaKazusa/S6MOD.

Dataset distillation has gained popularity as a technique for compressing large datasets into smaller, more efficient representations while retaining essential information for model training. Data features can be broadly divided into two types: instance-specific features, which capture unique, fine-grained details of individual examples, and class-general features, which represent shared, broad patterns across a class. However, previous approaches often struggle to balance these; some focus solely on class-general features, missing finer instance details, while others concentrate on instance-specific features, overlooking the shared characteristics essential for class-level understanding. In this paper, we propose the Non-Critical Region Refinement Dataset Distillation (NRR-DD) method, which preserves the instance-specific and fine-grained regions in synthetic data while enriching non-critical regions with more class-general information. This approach enables our models to leverage all pixel information to capture both types of features, thereby improving overall performance. Furthermore, we introduce Distance-Based Representative (DBR) knowledge transfer, which eliminates the need for soft labels in training by relying solely on the distance between synthetic data predictions and one-hot encoded labels. Experimental results demonstrate that our NRR-DD achieves state-of-the-art performance on both small-scale and large-scale datasets. Additionally, by storing only two distances per instance, our method achieves comparable results across various settings. Code will be available at https://github.com/tmtuan1307/NRR-DD.

Accurate 3D gaze estimation in unconstrained real-world environments remains a significant challenge due to variations in appearance, head pose, occlusion, and the limited availability of in-the-wild 3D gaze datasets. To address these challenges, we introduce a novel Self-Training Weakly-Supervised Gaze Estimation framework (ST-SWGE). This two-stage learning framework leverages diverse 2D gaze datasets, such as gaze-following data, which offer rich variations in appearances, natural scenes, and gaze distributions, and proposes an approach to generate 3D pseudo-labels and enhance model generalization. Furthermore, traditional modality-specific models, designed separately for images or videos, limit the effective use of available training data. To overcome this, we propose the Gaze Transformer (GaT), a modality-agnostic architecture capable of simultaneously learning static and dynamic gaze information from both image and video datasets. By combining 3D video datasets with 2D gaze target labels from gaze following tasks, our approach achieves the following key contributions: (i) Significant state-of-the-art improvements in within-domain and cross-domain generalization on unconstrained benchmarks like Gaze360 and GFIE, with notable cross-modal gains in video gaze estimation; (ii) Superior cross-domain performance on datasets such as MPIIFaceGaze and Gaze360 compared to frontal face methods. Code and pre-trained models will be released to the community.

Driven by the great success of Large Language Models (LLMs) in the 2D image domain, their applications in 3D scene understanding has emerged as a new trend. A key difference between 3D and 2D is that the situation of an egocentric observer in 3D scenes can change, resulting in different descriptions (e.g., "left" or "right"). However, current LLM-based methods overlook the egocentric perspective and simply use datasets from a global viewpoint. To address this issue, we propose a novel approach to automatically generate a situation-aware dataset by leveraging the scanning trajectory during data collection and utilizing Vision-Language Models (VLMs) to produce high-quality captions and question-answer pairs. Furthermore, we introduce a situation grounding module to explicitly predict the position and orientation of observer's viewpoint, thereby enabling LLMs to ground situation description in 3D scenes. We evaluate our approach on several benchmarks, demonstrating that our method effectively enhances the 3D situational awareness of LLMs while significantly expanding existing datasets and reducing manual effort.

The unprecedented advancements in Large Language Models (LLMs) have profoundly impacted natural language processing but have yet to fully embrace the realm of scalable vector graphics (SVG) generation. While LLMs encode partial knowledge of SVG data from web pages during training, recent findings suggest that semantically ambiguous and tokenized representations within LLMs may result in hallucinations in vector primitive predictions. Additionally, LLM training typically lacks modeling and understanding of the rendering sequence of vector paths, which can lead to occlusion between output vector primitives. In this paper, we present LLM4SVG, an initial yet substantial step toward bridging this gap by enabling LLMs to better understand and generate vector graphics. LLM4SVG facilitates a deeper understanding of SVG components through learnable semantic tokens, which precisely encode these tokens and their corresponding properties to generate semantically aligned SVG outputs. Using a series of learnable semantic tokens, a structured dataset for instruction following is developed to support comprehension and generation across two primary tasks. Our method introduces a modular architecture to existing large language models, integrating semantic tags, vector instruction encoders, fine-tuned commands, and powerful LLMs to tightly combine geometric, appearance, and language information. To overcome the scarcity of SVG-text instruction data, we developed an automated data generation pipeline that collected our SVGX-SFT Dataset, consisting of high-quality human-designed SVGs and 580k SVG instruction following data specifically crafted for LLM training, which facilitated the adoption of the supervised fine-tuning strategy popular in LLM development. By exploring various training strategies, we developed LLM4SVG, which significantly moves beyond optimized rendering-based approaches and language-model-based baselines to achieve remarkable results in human evaluation tasks. Code, model, and data will be released at: https://ximinng.github.io/LLM4SVGProject/