ICCV 2025 论文列表

We present a novel framework for training 3D image-conditioned diffusion models using only 2D supervision. Recovering 3D structure from 2D images is inherently ill-posed due to the ambiguity of possible reconstructions, making generative models a natural choice. However, most existing 3D generative models rely on full 3D supervision, which is impractical due to the scarcity of large-scale 3D datasets. To address this, we propose leveraging sparse-view supervision as a scalable alternative. While recent reconstruction models use sparse-view supervision with differentiable rendering to lift 2D images to 3D, they are predominantly deterministic, failing to capture the diverse set of plausible solutions and producing blurry predictions in uncertain regions. A key challenge in training 3D diffusion models with 2D supervision is that the standard training paradigm requires both the denoising process and supervision to be in the same modality. We address this by decoupling the noisy samples being denoised from the supervision signal, allowing the former to remain in 3D while the latter is provided in 2D. Our approach leverages suboptimal predictions from a deterministic image-to-3D model--acting as a "teacher"--to generate noisy 3D inputs, enabling effective 3D diffusion training without requiring full 3D ground truth. We validate our framework on both object-level and scene-level datasets, using two different 3D Gaussian Splat (3DGS) teachers. Our results show that our approach consistently improves upon these deterministic teachers, demonstrating its effectiveness in scalable and high-fidelity 3D generative modeling.

Visual Place Recognition (VPR) enables coarse localization by comparing query images to a reference database of geo-tagged images. Recent breakthroughs in deep learning architectures and training regimes have led to methods with improved robustness to factors like environment appearance change, but with the downside that the required training and/or matching compute scales with the number of distinct environmental conditions encountered. Here, we propose Hyperdimensional One Place Signatures (HOPS) to simultaneously improve the performance, compute and scalability of these state-of-the-art approaches by fusing the descriptors from multiple reference sets captured under different conditions. HOPS scales to any number of environmental conditions by leveraging the Hyperdimensional Computing framework. Extensive evaluations demonstrate that our approach is highly generalizable and consistently improves recall performance across all evaluated VPR methods and datasets by large margins. Arbitrarily fusing reference images without compute penalty enables numerous other useful possibilities, three of which we demonstrate here: improved performance with reduced dimensionality descriptors, stacking synthetic images, and coarse localization to an entire traverse or environmental section.

Generalized Category Discovery (GCD) aims to classify unlabeled data from both known and unknown categories by leveraging knowledge from labeled known categories. While existing methods have made notable progress, they often overlook a hidden stumbling block in GCD: distracted attention. Specifically, when processing unlabeled data, models tend to focus not only on key objects in the image but also on task-irrelevant background regions, leading to suboptimal feature extraction. To remove this stumbling block, we propose Attention Focusing (AF), an adaptive mechanism designed to sharpen the model's focus by pruning non-informative tokens. AF consists of two simple yet effective components: Token Importance Measurement (TIME) and Token Adaptive Pruning (TAP), working in a cascade. TIME quantifies token importance across multiple scales, while TAP prunes non-informative tokens by utilizing the multi-scale importance scores provided by TIME. AF is a lightweight, plug-and-play module that integrates seamlessly into existing GCD methods with minimal computational overhead. When incorporated into one prominent GCD method, SimGCD, AF achieves up to 15.4% performance improvement over the baseline with minimal computational overhead. The implementation code is provided in https://github.com/Afleve/AFGCD.

Knowledge distillation (KD) is an effective method for enhancing a small model, named student, by training it under the supervision of larger teacher models. However, existing studies indicate that a substantial capacity gap between the student and teacher can lead to poor learning for the student model. This capacity gap problem limits the applicability of KD and necessitates careful selection of the teacher's size. Despite its importance, the underlying cause of the capacity gap problem remains underexplored. In this paper, we reveal that a substantial disparity in the output distributions of teacher and student models is a key factor behind this issue. To demonstrate this, we decompose the KD loss into two components: class-wise similarity and intra-class distribution, and analyze the contribution of each term. Our analysis shows that a large distributional mismatch can lead to poor student learning.Inspired by this observation, we propose the Adapted Intra-class Distribution (AID) method, wherein the teacher model is fine-tuned to optimize its intra-class distribution to better align with the student's capacity prior to knowledge distillation. This approach effectively bridges the capacity gap between teacher and student models and consistently achieves state-of-the-art performance across a diverse range of architectures.

Foundation models (FMs) excel in zero-shot tasks but benefit from task-specific adaptation. However, privacy concerns prevent data sharing among multiple data owners, and proprietary restrictions prevent the learning service provider (LSP) from sharing the FM. In this work, we propose BlindFed, a framework enabling collaborative FM adaptation while protecting both parties: data owners do not access the FM or each other's data, and the LSP does not see sensitive task data. BlindFed relies on fully homomorphic encryption (FHE) and consists of three key innovations: (i) FHE-friendly architectural modifications via polynomial approximations and low-rank adapters, (ii) a two-stage split learning approach combining offline knowledge distillation and online encrypted inference for adapter training without backpropagation through the FM, and (iii) a privacy-boosting scheme using sample permutations and stochastic block sampling to mitigate model extraction attacks. Empirical results on four image classification datasets demonstrate the practical feasibility of the BlindFed framework, albeit at a high communication cost and large computational complexity for the LSP. The code can be found at https://github.com/tnurbek/blindfed.

End-to-end optimization, which integrates differentiable optics simulators with computational algorithms, enables the joint design of hardware and software for data-driven imaging systems. However, existing methods usually compromise physical accuracy by neglecting wave optics or off-axis effects due to the high computational cost of modeling both aberration and diffraction. This limitation raises concerns about the robustness of optimized designs. In this paper, we propose a differentiable optics simulator that accurately and efficiently models aberration and diffraction in compound optics and allows us to analyze the role and impact of diffraction in end-to-end optimization. Experimental results demonstrate that compared with ray-optics-based optimization, diffraction-aware optimization improves system robustness to diffraction blur. Through accurate wave optics modeling, we also apply the simulator to optimize the Fizeau interferometer and freeform optics elements. These findings underscore the importance of accurate wave optics modeling in robust end-to-end optimization. Our code is publicly available at: https://github.com/JerryHoTaiwan/DeepWaveOptics

3D Gaussian Splatting (3DGS) is a powerful reconstruction technique; however, it requires initialization from accurate camera poses and high-fidelity point clouds. Typically, the initialization is taken from Structure-from-Motion (SfM) algorithms; however, SfM is time-consuming and restricts the application of 3DGS in real-world scenarios and large-scale scene reconstruction. We introduce a constrained optimization method for simultaneous camera pose estimation and 3D reconstruction that does not require SfM support. Core to our approach is decomposing a camera pose into a sequence of camera-to-(device-)center and (device-)center-to-world optimizations. To facilitate, we propose two optimization constraints conditioned to the sensitivity of each parameter group and restricts each parameter's search space. In addition, as we learn the scene geometry directly from the noisy point clouds, we propose geometric constraints to improve the reconstruction quality. Experiments demonstrate that the proposed method significantly outperforms the existing (multi-modal) 3DGS baseline and methods supplemented by COLMAP on both our collected dataset and two public benchmarks.

Compositional Zero-Shot Learning (CZSL) aims to recognize unseen combinations of known objects and attributes by leveraging knowledge from previously seen compositions. Traditional approaches primarily focus on disentangling attributes and objects, treating them as independent entities during learning. However, this assumption overlooks the semantic constraints and contextual dependencies inside a composition. For example, certain attributes naturally pair with specific objects (e.g., "striped" applies to "zebra" or "shirts" but not "sky" or "water"), while the same attribute can manifest differently depending on context (e.g., "young" in "young tree" *vs* "young dog"). Thus, capturing attribute-object interdependence remains a fundamental yet long-ignored challenge in CZSL.In this paper, we adopt a Conditional Probability Framework (CPF) to explicitly model attribute-object dependencies. We decompose the probability of a composition into two components: the likelihood of an object and the conditional likelihood of its attribute. To enhance object feature learning, we incorporate textual descriptors to highlight semantically relevant image regions. These enhanced object features then guide attribute learning through a cross-attention mechanism, ensuring better contextual alignment. By jointly optimizing object likelihood and conditional attribute likelihood, our method effectively captures compositional dependencies and generalizes well to unseen compositions. Extensive experiments on multiple CZSL benchmarks demonstrate the superiority of our approach. Code is available at here.

Real-time rendering of dynamic scenes with view-dependent effects remains a fundamental challenge in computer graphics. While recent advances in Gaussian Splatting have shown promising results separately handling dynamic scenes (4DGS) and view-dependent effects (6DGS), no existing method unifies these capabilities while maintaining real-time performance. We present 7D Gaussian Splatting (7DGS), a unified framework representing scene elements as seven-dimensional Gaussians spanning position (3D), time (1D), and viewing direction (3D). Our key contribution is an efficient conditional slicing mechanism that transforms 7D Gaussians into view- and time-conditioned 3D Gaussians, maintaining compatibility with existing 3D Gaussian Splatting pipelines while enabling joint optimization. Experiments demonstrate that 7DGS outperforms prior methods by up to 7.36 dB in PSNR while achieving real-time rendering (401 FPS) on challenging dynamic scenes with complex view-dependent effects.

Efficient and accurate object pose estimation is an essential component for modern vision systems in many applications such as Augmented Reality, autonomous driving, and robotics. While research in model-based 6D object pose estimation has delivered promising results, model-free methods are hindered by the high computational load in rendering and inferring consistent poses of arbitrary objects in a live RGB-D video stream. To address this issue, we present 6DOPE-GS, a novel method for online 6D object pose estimation & tracking with a single RGB-D camera by effectively leveraging advances in Gaussian Splatting. Thanks to the fast differentiable rendering capabilities of Gaussian Splatting, 6DOPE-GS can simultaneously optimize for 6D object poses and 3D object reconstruction. To achieve the necessary efficiency and accuracy for live tracking, our method uses incremental 2D Gaussian Splatting with an intelligent dynamic keyframe selection procedure to achieve high spatial object coverage and prevent erroneous pose updates. We also propose an opacity statistic-based pruning mechanism for adaptive Gaussian density control, to ensure training stability and efficiency. We evaluate our method on the HO3D and YCBInEOAT datasets and show that 6DOPE-GS matches the performance of state-of-the-art baselines for model-free simultaneous 6D pose tracking and reconstruction while providing a 5x speedup. We also demonstrate the method's suitability for live, dynamic object tracking and reconstruction in a real-world setting.

4D panoptic segmentation in a streaming setting is critical for highly dynamic environments, such as evacuating dense crowds and autonomous driving in complex scenarios, where real-time, fine-grained perception within a constrained time budget is essential. In this paper, we introduce 4DSegStreamer, a novel framework that employs a Dual-Thread System to efficiently process streaming frames. The framework is general and can be seamlessly integrated into existing 3D and 4D segmentation methods to enable real-time capability. It also demonstrates superior robustness compared to existing streaming perception approaches, particularly under high FPS conditions. The system consists of a predictive thread and an inference thread. The predictive thread leverages historical motion and geometric information to extract features and forecast future dynamics. The inference thread ensures timely prediction for incoming frames by aligning with the latest memory and compensating for ego-motion and dynamic object movements. We evaluate 4DSegStreamer on the indoor HOI4D dataset and the outdoor SemanticKITTI and nuScenes datasets. Comprehensive experiments demonstrate the effectiveness of our approach, particularly in accurately predicting dynamic objects in complex scenes.

Multimodal Large Language Models (MLLMs) have demonstrated impressive 2D image/video understanding capabilities.However, there are no publicly standardized benchmarks to assess the abilities of MLLMs in understanding the 4D objects.In this paper, we introduce 4D-Bench, the first benchmark to evaluate the capabilities of MLLMs in 4D object understanding, featuring tasks in 4D object Question Answering (4D object QA) and 4D object captioning.4D-Bench provides 4D objects with diverse categories, high-quality annotations, and tasks necessitating multi-view spatial-temporal understanding, different from existing 2D image/video-based benchmarks.With 4D-Bench, we evaluate a wide range of open-source and closed-source MLLMs.The results from the 4D object captioning experiment indicate that MLLMs generally exhibit weaker temporal understanding compared to their appearance understanding, notably, while open-source models approach closed-source performance in appearance understanding, they show larger performance gaps in temporal understanding.4D object QA yields surprising findings: even with simple single-object videos, MLLMs perform poorly, with state-of-the-art GPT-4o achieving only 63% accuracy compared to the human baseline of 91%.These findings highlight a substantial gap in 4D object understanding and the need for further advancements in MLLMs.

General visual representations learned from web-scale datasets for robotics have achieved great success in recent years, enabling data-efficient robot learning on manipulation tasks; yet these pre-trained representations are mostly on 2D images, neglecting the inherent 3D nature of the world. However, due to the scarcity of large-scale 3D data, it is still hard to extract a universal 3D representation from web datasets. Instead, we are seeking a general visual pre-training framework that could improve all 3D representations as an alternative. Our framework, called FVP, is a novel 4D Visual Pre-training framework for real-world robot learning. FVP frames the visual pre-training objective as a next-point-cloud-prediction problem, models the prediction model as a diffusion model, and pre-trains the model on the larger public datasets directly. Across twelve real-world manipulation tasks, FVP boosts the average success rate of 3D Diffusion Policy (DP3) for these tasks by 28%. The FVP pre-trained DP3 achieves state-of-the-art performance across imitation learning methods. Moreover, the efficacy of \ours adapts across various point cloud encoders and datasets. Finally, we apply FVP to the RDT-1B, a larger Vision-Language-Action robotic model, enhancing its performance on various robot tasks. Our project page is available at: https://4d-visual-pretraining.github.io/.

Simultaneously localizing camera poses and constructing Gaussian radiance fields in dynamic scenes establish a crucial bridge between 2D images and the 4D real world. Instead of removing dynamic objects as distractors and reconstructing only static environments, this paper proposes an efficient architecture that incrementally tracks camera poses and establishes the 4D Gaussian radiance fields in unknown scenarios by using a sequence of RGB-D images. First, by generating motion masks, we obtain static and dynamic priors for each pixel. To eliminate the influence of static scenes and improve the efficiency of learning the motion of dynamic objects, we classify the Gaussian primitives into static and dynamic Gaussian sets, while the sparse control points along with an MLP are utilized to model the transformation fields of the dynamic Gaussians. To more accurately learn the motion of dynamic Gaussians, a novel 2D optical flow map reconstruction algorithm is designed to render optical flows of dynamic objects between neighbor images, which are further used to supervise the 4D Gaussian radiance fields along with traditional photometric and geometric constraints. In experiments, qualitative and quantitative evaluation results show that the proposed method achieves robust tracking and high-quality view synthesis performance in real-world environments.

3D spatial reasoning is the ability to analyze and interpret the positions, orientations, and spatial relationships of objects within the 3D space. This allows models to develop a comprehensive understanding of the 3D scene, enabling their applicability to a broader range of applications, such as autonomous navigation, robotics, and AR/VR. Despite the remarkable improvements achieved by large multi-modal models (LMMs) in a wide range of image and video understanding tasks, their abilities to perform 3D spatial reasoning are less studied. In this work we present the first comprehensive 3D spatial reasoning benchmark, 3DSRBench, with 3,000 annotated image question answering triplets from 12 question types. We balance the data distribution by collecting complimentary images that lead to opposite answers given the same question. We also adopt a novel FlipEval for robust evaluation of 3D spatial reasoning capabilities. Moreover, to study the robustness of 3D spatial reasoning w.r.t. camera 3D viewpoints, our 3DSRBench involves two subsets with 3D spatial reasoning questions on images from the same scene with common and uncommon viewpoints. We benchmark a wide range of open-sourced and proprietary LMMs, revealing their limitations in different types of 3D awareness, i.e., height, orientation, location, and multi-object reasoning. Our 3DSRBench also allows us to study the design choices of developing LMMs with strong 3D reasoning capabilities, such as the vision encoders, connectors, and training recipes.

3D cars are widely used in self-driving systems, virtual and augmented reality, and gaming applications. However, existing 3D car datasets are either synthetic or low-quality, limiting their practical utility and leaving a significant gap with the high-quality real-world 3D car dataset. In this paper, we present the first large-scale 3D real car dataset, termed 3DRealCar, which offers three key features: (1) High-Volume: 2,500 cars meticulously scanned using smartphones to capture RGB images and point clouds with real-world dimensions; (2) High-Quality: Each car is represented by an average of 200 dense, high-resolution 360-degree RGB-D views, enabling high-fidelity 3D reconstruction; (3) High-Diversity: The dataset encompasses a diverse collection of cars from over 100 brands, captured under three distinct lighting conditions (reflective, standard, and dark). We further provide detailed car parsing maps for each instance to facilitate research in automotive segmentation tasks. To focus on vehicles, background point clouds are removed, and all cars are aligned to a unified coordinate system, enabling controlled reconstruction and rendering. We benchmark state-of-the-art 3D reconstruction methods across different lighting conditions using 3DRealCar. Extensive experiments demonstrate that the standard lighting subset can be used to reconstruct high-quality 3D car models that significantly enhance performance on various car-related 2D and 3D tasks. Notably, our dataset reveals critical challenges faced by current 3D reconstruction methods under reflective and dark lighting conditions, providing valuable insights for future research.

A 3D scene graph represents a compact scene model by capturing both the objects present and the semantic relationships between them, making it a promising structure for robotic applications. To effectively interact with users, an embodied intelligent agent should be able to answer a wide range of natural language queries about the surrounding 3D environment. Large Language Models (LLMs) are beneficial solutions for user-robot interaction due to their natural language understanding and reasoning abilities. Recent methods for learning scene representations have shown that adapting these representations to the 3D world can significantly improve the quality of LLM responses. However, existing methods typically rely only on geometric information, such as object coordinates, and overlook the rich semantic relationships between objects. In this work, we propose 3DGraphLLM, a method for constructing a learnable representation of a 3D scene graph that explicitly incorporates semantic relationships. This representation is used as input to LLMs for performing 3D vision-language tasks. In our experiments on popular ScanRefer, Multi3DRefer, ScanQA, Sqa3D, and Scan2cap datasets, we demonstrate that our approach outperforms baselines that do not leverage semantic relationships between objects. The code is publicly available at https://github.com/CognitiveAISystems/3DGraphLLM.

We present 3DGS-LM, a new method that accelerates the reconstruction of 3D Gaussian Splatting (3DGS) by replacing its ADAM optimizer with a tailored Levenberg-Marquardt (LM). Existing methods reduce the optimization time by decreasing the number of Gaussians or by improving the implementation of the differentiable rasterizer. However, they still rely on the ADAM optimizer to fit Gaussian parameters of a scene in thousands of iterations, which can take up to an hour. To this end, we change the optimizer to LM that runs in conjunction with the 3DGS differentiable rasterizer. For efficient GPU parallelization, we propose a caching data structure for intermediate gradients that allows us to efficiently calculate Jacobian-vector products in custom CUDA kernels. In every LM iteration, we calculate update directions from multiple image subsets using these kernels and combine them in a weighted mean. Overall, our method is 20% faster than the original 3DGS while obtaining the same reconstruction quality. Our optimization is also agnostic to other methods that accelerate 3DGS, thus enabling even faster speedups compared to vanilla 3DGS.

Monocular 3D object detection is valuable for various applications such as robotics and AR/VR. Existing methods are confined to closed-set settings, where the training and testing sets consist of the same scenes and/or object categories. However, real-world applications often introduce new environments and novel object categories, posing a challenge to these methods. In this paper, we address monocular 3D object detection in an open-set setting and introduce the first end-to-end 3D Monocular Open-set Object Detector (3D-MOOD). We propose to lift the open-set 2D detection into 3D space through our designed 3D bounding box head, enabling end-to-end joint training for both 2D and 3D tasks to yield better overall performance. We condition the object queries with geometry prior and overcome the generalization for 3D estimation across diverse scenes. To further improve performance, we design the canonical image space for more efficient cross-dataset training. We evaluate 3D-MOOD on both closed-set settings (Omni3D) and open-set settings (Omni3D - Argoverse 2, ScanNet), and achieve new state-of-the-art results. Code and models are available at royyang0714.github.io/3D-MOOD.