3DV 2025 论文列表

We present Rig3DGS, a novel technique for creating reanimatable 3D portraits from short monocular smartphone videos. Rig3DGS learns to reconstruct a set of controllable 3D gaussians from a monocular video of a dynamic subject captured with varying head poses and facial expressions in an in-the-wild scene. In contrast to synchronized multi-view studio captures, this in-the-wild, single camera setup brings fresh challenges to learning high quality 3D gaussians. We address these challenges by learning to deform 3D gaussians from a fixed canonical space to the deformed space that is consistent with the target facial expression and head-pose. Our key contribution is a carefully designed deformation model that is guided by a 3D face morphable model. This deformation not only enables control over facial expression and head-poses but also allows our method to generates high quality photorealistic renders of the whole scene. Once trained, Rig3DGS is able to generate photorealistic renders of a subject and their scene for novel facial expression, head-poses, and viewing directions. Through extensive experiments we demonstrate that Rig3DGS significantly outperforms prior art while being orders of magnitude faster.

Cross-modal place recognition methods are flexible GPS-alternatives under varying environment conditions and sensor setups. However, this task is non-trivial since extracting consistent and robust global descriptors from different modalities is challenging. To tackle this issue, we propose \textit{Voxel-Cross-Pixel (VXP)}, a novel camera-to-LiDAR place recognition framework that enforces local similarities in a self-supervised manner and effectively brings global context from images and LiDAR scans into a shared feature space. Specifically, VXP is trained in three stages: first, we deploy a visual transformer to compactly represent input images. Secondly, we establish local correspondences between image-based and point cloud-based feature spaces using our novel geometric alignment module. We then aggregate local similarities into an expressive shared latent space. Extensive experiments on the three benchmarks (Oxford RobotCar, ViViD++ and KITTI) demonstrate that our method surpasses the state-of-the-art cross-modal retrieval by a large margin. Our evaluations show that the proposed method is accurate, efficient and light-weight. Our project page is available at: \href{https://yunjinli.github.io/projects-vxp/}{https://yunjinli.github.io/projects-vxp/}.

3D Gaussian Splatting (3DGS) has attracted great attention in novel view synthesis because of its superior rendering efficiency and high fidelity. However, the trained Gaussians suffer from severe zooming degradation due to non-adjustable representation derived from single-scale training. Though some methods attempt to tackle this problem via post-processing techniques such as selective rendering or filtering techniques towards primitives, the scale-specific information is not involved in Gaussians. In this paper, we propose a unified optimization method to make Gaussians adaptive for arbitrary scales by self-adjusting the primitive properties (e.g., color, shape and size) and distribution (e.g., position). Inspired by the mipmap technique, we design pseudo ground-truth for the target scale and propose a scale-consistency guidance loss to inject scale information into 3D Gaussians. Our method is a plug-in module, applicable for any 3DGS models to solve the zoom-in and zoom-out aliasing. Extensive experiments demonstrate the effectiveness of our method. Notably, our method outperforms 3DGS in PSNR by an average of 9.25 dB for zoom-in and 10.40 dB for zoom-out on the NeRF Synthetic dataset.

We introduce Spurfies, a novel method for sparse-view surface reconstruction that disentangles appearance and geometry information to utilize local geometry priors trained on synthetic data. Recent research heavily focuses on 3D reconstruction using dense multi-view setups, typically requiring hundreds of images. However, these methods often struggle with few-view scenarios. Existing sparse-view reconstruction techniques often rely on multi-view stereo networks that need to learn joint priors for geometry and appearance from a large amount of data. In contrast, we introduce a neural point representation that disentangles geometry and appearance to train a local geometry prior using a subset of the synthetic ShapeNet dataset only. During inference, we utilize this surface prior as additional constraint for surface and appearance reconstruction from sparse input views via differentiable volume rendering, restricting the space of possible solutions. We validate the effectiveness of our method on the DTU dataset and demonstrate that it outperforms previous work by 35\% in surface quality while achieving competitive novel view synthesis quality. Moreover, in contrast to previous works, our method can be applied to larger, unbounded scenes, such as Mip-NeRF360.

3D Gaussian Splatting (3DGS) has recently enabled real-time rendering of unbounded 3D scenes for novel view synthesis. However, this technique requires dense training views to accurately reconstruct 3D geometry. A limited number of input views will significantly degrade reconstruction quality, resulting in artifacts such as "floaters'' and "background collapse'' at unseen viewpoints. In this work, we introduce SparseGS, an efficient training pipeline designed to address the limitations of 3DGS in scenarios with sparse training views. SparseGS incorporates depth priors, novel depth rendering techniques, and a pruning heuristic to mitigate floater artifacts, alongside an Unseen Viewpoint Regularization module to alleviate background collapses. Our extensive evaluations on the Mip-NeRF360, LLFF, and DTU datasets demonstrate that SparseGS achieves high-quality reconstruction in both unbounded and forward-facing scenarios, with as few as 12 and 3 input images, respectively, while maintaining fast training and real-time rendering capabilities.

We present Spann3R, a novel approach for dense 3D reconstruction from ordered or unordered image collections. Built on the DUSt3R paradigm, Spann3R uses a transformer-based architecture to directly regress a global pointmap from images without any prior knowledge of the scene or camera parameters. Unlike DUSt3R, which predicts per image-pair pointmaps each expressed in its local coordinate frame, Spann3R can predict a per-image pointmaps expressed in a global coordinate system, thus eliminating the need for optimization-based global alignment. The key idea of Spann3R is to manage an external spatial memory that learns to keep track of all previous relevant 3D information. Spann3R then queries this spatial memory to predict the 3D structure of the next frame in a global coordinate system. Taking advantage of DUSt3R's pre-trained weights, and further fine-tuning on a subset of datasets, Spann3R shows competitive performance and generalization ability on various unseen datasets and is able to process ordered image collections in real-time.

To evaluate end-to-end autonomous driving systems, a simulation environment based on Novel View Synthesis (NVS) techniques is essential, which synthesizes photo-realistic images and point clouds from previously recorded sequences under new vehicle poses, particularly in cross-lane scenarios. Therefore, the development of a multi-lane dataset and benchmark is necessary. While recent synthetic scene-based NVS datasets have been prepared for cross-lane benchmarking, they still lack the realism of captured images and point clouds. To further assess the performance of existing methods based on NeRF and 3DGS, we present the first multi-lane dataset registering parallel scans specifically for novel driving view synthesis dataset derived from real-world scans, comprising 25 groups of associated sequences, including 16,000 front-view images, 64,000 surround-view images, and 16,000 LiDAR frames. All frames are labeled to differentiate moving objects from static elements. Using this dataset, we evaluate the performance of existing approaches in various testing scenarios at different lanes and distances. Additionally, our method provides the solution for solving and assessing the quality of multi-sensor poses for multi-modal data alignment for curating such a dataset in real-world. We plan to continually add new sequences to test the generalization of existing methods across different scenarios. The dataset is released publicly at the project page: https://nizqleo.github.io/paralane-dataset/.

Test-Time Training (TTT) proposes to adapt a pre-trained network to changing data distributions on-the-fly. In this work, we propose the first TTT method for 3D semantic segmentation, TTT-KD, which models Knowledge Distillation (KD) from foundation models (e.g. DINOv2) as a self-supervised objective for adaptation to distribution shifts at test-time. Given access to paired image-pointcloud (2D-3D) data, we first optimize a 3D segmentation backbone for the main task of semantic segmentation using the pointclouds and the task of 2D → 3D KD by using an off-the-shelf 2D pre-trained foundation model. At test-time, our TTT-KD updates the 3D segmentation backbone for each test sample, by using the self-supervised task of knowledge distillation, before performing the final prediction. Extensive evaluations on multiple indoor and outdoor 3D segmentation benchmarks show the utility of TTT-KD, as it improves performance for both in-distribution (ID) and out-of-distribution (OOD) test datasets. We achieve a gain of up to 13 % mIoU (7 % on average) when the train and test distributions are similar and up to 45 % (20 % on average) when adapting to OOD test samples

Implicit neural representations (INRs) have become an important method for encoding various data types, such as 3D objects or scenes, images, and videos. They have proven to be particularly effective at representing 3D content, e.g., 3D scene reconstruction from 2D images, novel 3D content creation, as well as the representation, interpolation, and completion of 3D shapes. With the widespread generation of 3D data in an INR format, there is a need to support effective organization and retrieval of INRs saved in a data store. A key aspect of retrieval and clustering of INRs in a data store is the formulation of similarity between INRs that would, for example, enable retrieval of similar INRs using a query INR. In this work, we propose INRet, a method for determining similarity between INRs that represent shapes, thus enabling accurate retrieval of similar shape INRs from an INR data store. INRet flexibly supports different INR architectures such as INRs with octree grids, triplanes, and hash grids, as well as different implicit functions including signed/unsigned distance function and occupancy field. We demonstrate that our method is more general and accurate than the existing INR retrieval method, which only supports simple MLP INRs and requires the same architecture between the query and stored INRs. Furthermore, compared to converting INRs to other representations (e.g., point clouds or multi-view images) for 3D shape retrieval, INRet achieves higher accuracy while avoiding the conversion overhead.

We present a method to generate 3D objects in styles. Our method takes a text prompt and a style reference image as input and reconstructs a neural radiance field to synthesize a 3D model with the content aligning with the text prompt and the style following the reference image. To simultaneously generate the 3D object and perform style transfer in one go, we propose a stylized score distillation loss to guide a text-to-3D optimization process to output visually plausible geometry and appearance. Our stylized score distillation is based on a combination of an original pretrained text-to-image model and its modified sibling with the key and value features of self-attention layers manipulated to inject styles from the reference image. Comparisons with state-of-the-art methods demonstrated the strong visual performance of our method, further supported by the quantitative results from our user study.

Building a foundation model for 3D vision is a complex challenge that remains unsolved. Towards that goal, it is important to understand the 3D reasoning capabilities of current models as well as identify the gaps between these models and humans. Therefore, we construct a new 3D visual understanding benchmark named UniQA-3D. UniQA-3D covers fundamental 3D vision tasks in the Visual Question Answering (VQA) format. We evaluate state-of-the-art Vision-Language Models (VLMs), specialized models, and human subjects on it. Our results show that VLMs generally perform poorly, while the specialized models are accurate but not robust, failing under geometric perturbations. In contrast, human vision continues to be the most reliable 3D visual system. We further demonstrate that neural networks align more closely with human 3D vision mechanisms compared to classical computer vision methods, and Transformer-based networks such as ViT align more closely with human 3D vision mechanisms than CNNs. We hope our study will benefit the future development of foundation models for 3D vision. Code is available at https://github.com/princeton-vl/UniQA-3D.

With the attention gained by camera-only 3D object detection in autonomous driving, methods based on Bird-Eye-View (BEV) representation especially derived from the forward view transformation paradigm, i.e., lift-splat-shoot (LSS), have recently seen significant progress. The BEV representation formulated by the frustum based on depth distribution prediction is ideal for learning the road structure and scene layout from multi-view images. However, to retain computational efficiency, the compressed BEV representation such as in resolution and axis is inevitably weak in retaining the individual geometric details, undermining the methodological generality and applicability. With this in mind, to compensate for the missing details and utilize multi-view geometry constraints, we propose LSSInst, a two-stage object detector incorporating BEV and instance representations in tandem. The proposed detector exploits fine-grained pixel-level features that can be flexibly integrated into existing LSS-based BEV networks. Having said that, due to the inherent gap between two representation spaces, we design the instance adaptor for the BEV-to-instance semantic coherence rather than pass the proposal naively. Extensive experiments demonstrated that our proposed framework is of excellent generalization ability and performance, which boosts the performances of modern LSS-based BEV perception methods without bells and whistles and outperforms current LSS-based state-of-the-art works on the large-scale nuScenes benchmark.

Mesh extraction from occupancy functions is a useful tool in creating synthetic datasets for computer vision. However, existing mesh extraction methods have artifacts or performance profiles that limit their use. We propose OcMesher, a mesh extractor that efficiently handles high-detail unbounded scenes with perfect view consistency, with easy export to downstream real-time engines. The main novelty of our solution is an algorithm to construct an octree based on a given occupancy function and multiple camera views. We performed extensive experiments, and demonstrate OcMesher's usefulness for synthetic training & benchmark datasets, generating real-time environments for embodied AI and mesh extraction from depthmaps or novel view synthesis methods.

Modern visual effects (VFX) software has made it possible for skilled artists to create imagery of virtually anything. However, the creation process remains laborious, complex, and largely inaccessible to everyday users. In this work, we present AutoVFX, a framework that automatically creates realistic and dynamic VFX videos from a single video and natural language instructions. By carefully integrating neural scene modeling, LLM-based code generation, and physical simulation, AutoVFX is able to provide physically-grounded, photorealistic editing effects that can be controlled directly using natural language instructions. We conduct extensive experiments to validate AutoVFX's efficacy across a diverse spectrum of videos and instructions. Quantitative and qualitative results suggest that AutoVFX outperforms all competing methods by a large margin in generative quality, instruction alignment, editing versatility, and physical plausibility.

This paper addresses the problem of reconstructing a scene online at the level of objects given an RGB-D video sequence. While current object-aware neural implicit representations hold promise, they are limited in online reconstruction efficiency and shape completion. Our main contributions to alleviate the above limitations are twofold. First, we propose a feature grid interpolation mechanism to continuously update grid-based object-centric neural implicit representations as new object parts are revealed. Second, we construct an object library with previously mapped objects in advance and leverage the corresponding shape priors to initialize geometric object models in new videos, subsequently completing them with novel views as well as synthesized past views to avoid losing original object details. Extensive experiments on synthetic environments from the Replica dataset, real-world ScanNet sequences and videos captured in our laboratory demonstrate that our approach outperforms state-of-the-art neural implicit models for this task in terms of reconstruction accuracy and completeness.

We present a method to reconstruct time-consistent human body models from monocular videos, focusing on extremely loose clothing or handheld object interactions. Prior work in human reconstruction is either limited to tight-fitting clothing with no object interactions, or requires calibrated multi-view captures or personalized template scans which are costly to collect at scale. Our key insight for high-quality yet flexible reconstruction is the careful combination of generic human priors about articulated body shape (learned from large-scale training data) with video-specific articulated "bag-of-bones" deformation (fit to a single video via test-time optimization). We accomplish this by learning a neural implicit model that disentangles body versus clothing deformations as separate motion model layers. To capture subtle geometry of clothing, we leverage image-based priors such as human body pose, surface normals, and optical flow during optimization. The resulting neural fields can be extracted into time-consistent meshes, or further refined as explicit 3D gaussians for high-fidelity interactive rendering. On datasets with highly challenging clothing deformations and object interactions, DressRecon yields higher-fidelity 3D reconstructions than prior art.

Synthesizing 3D whole bodies that realistically grasp objects is useful for animation, mixed reality, and robotics. This is challenging, because the hands and body need to look natural w.r.t. each other, the grasped object, as well as the local scene (i.e., a receptacle supporting the object). Moreover, training data for this task is really scarce, while capturing new data is expensive. Recent work goes beyond finite datasets via a divide-and-conquer approach; it first generates a “guiding” right-hand grasp, and then searches for bodies that match this. However, the guiding-hand synthesis lacks controllability and receptacle awareness, so it likely has an implausible direction (i.e., a body can’t match this without penetrating the receptacle) and needs corrections through major post-processing. Moreover, the body search needs exhaustive sampling and is expensive. These are strong limitations. We tackle these with a novel method called CWGrasp. Our key idea is that performing geometry-based reasoning “early on,” instead of “too late,” provides rich “control” signals for inference. To this end, CWGrasp first samples a plausible reaching-direction vector (used later for both the arm and hand) from a probabilistic model built via ray-casting from the object and collision checking. Then, it generates a reaching body with a desired arm direction, as well as a “guiding” grasping hand with a desired palm direction that complies with the arm’s one. Eventually, CWGrasp refines the body to match the “guiding” hand, while plausibly contacting the scene. Notably, generating already-compatible “parts” greatly simplifies the “whole”. Moreover, CWGrasp uniquely tackles both right and left-hand grasps. We evaluate on the GRAB and ReplicaGrasp datasets. CWGrasp outperforms baselines, at lower runtime and budget, while all components help performance. Code and models are available at https://gpaschalidis.github.io/cwgrasp.

The 3D occupancy estimation task has become an important challenge in the area of vision-based autonomous driving recently. However, most existing camera-based methods rely on costly 3D voxel labels or LiDAR scans for training, limiting their practicality and scalability. Moreover, most methods are tied to a predefined set of classes which they can detect. In this work we present a novel approach for open vocabulary occupancy estimation called LangOcc, that is trained only via camera images, and can detect arbitrary semantics via vision-language alignment. In particular, we distill the knowledge of the strong vision-language aligned encoder CLIP into a 3D occupancy model via differentiable volume rendering. Our model estimates vision-language aligned features in a 3D voxel grid using only images. It is trained in a weakly-supervised manner by rendering our estimations back to 2D space, where features can easily be aligned with CLIP. This training mechanism automatically supervises the scene geometry, allowing for a straight-forward and powerful training method without any explicit geometry supervision. LangOcc outperforms LiDAR-supervised competitors in open vocabulary occupancy with a mAP of $22.7$ by a large margin ($+4.3 \%$), solely relying on vision-based training. We also achieve a mIoU score of $11.84$ on the Occ3D-nuScenes dataset, surpassing previous vision-only semantic occupancy estimation methods ($+1.71\%$), despite not being limited to a specific set of categories.

We introduce GaussianAvatar-Editor, an innovative framework for text-driven editing of animatable Gaussian head avatars that can be fully controlled in expression, pose, and viewpoint. Unlike static 3D Gaussian editing, editing animatable 4D Gaussian avatars presents challenges related to motion occlusion and spatial-temporal inconsistency. To address these issues, we propose the Weighted Alpha Blending Equation (WABE). This function enhances the blending weight of visible Gaussians while suppressing the influence on non-visible Gaussians, effectively handling motion occlusion during editing. Furthermore, to improve editing quality and ensure 4D consistency, we incorporate conditional adversarial learning into the editing process. This strategy helps to refine the edited results and maintain consistency throughout the animation. By integrating these methods, our GaussianAvatar-Editor achieves photorealistic and consistent results in animatable 4D Gaussian editing. We conduct comprehensive experiments across various subjects to validate the effectiveness of our proposed techniques, which demonstrates the superiority of our approach over existing methods. More results and code are available at: \url{https://xiangyueliu.github.io/GaussianAvatar-Editor/}.