3DV 2025 Papers

Novel view synthesis techniques predominantly utilize RGB cameras, inheriting their limitations such as the need for sufficient lighting, susceptibility to motion blur, and restricted dynamic range. In contrast, event cameras are significantly more resilient to these limitations but have been less explored in this domain, particularly in large-scale settings. Current methodologies primarily focus on front-facing or object-oriented (360-degree view) scenarios. For the first time, we introduce 3D Gaussians for event-based novel view synthesis. Our method reconstructs large and unbounded scenes with high visual quality. We contribute the first real and synthetic event datasets tailored for this setting. Our method demonstrates superior novel view synthesis and consistently outperforms the baseline EventNeRF by a margin of 11−25% in PSNR (dB) while being orders of magnitude faster in reconstruction and rendering.

Separable 3D reconstruction of multiple objects from multi-view RGB images—resulting in two different 3D shapes for the two objects with a clear separation between them—remains a sparely researched problem. It is challenging due to severe mutual occlusions and ambiguities along the objects’ interaction boundaries. This paper investigates the setting and introduces a new neuro-implicit method that can reconstruct the geometry and appearance of two objects undergoing close interactions while disjoining both in 3D, avoiding surface inter-penetrations and enabling novel-view synthesis of the observed scene. In our approach, the objects in the scene are first encoded using a shared multi-resolution hash grid. Next, its features are decoded into two neural SDFs for the respective objects. The framework is end-to-end trainable and supervised using a novel alpha-blending regularization that ensures that the two geometries are well separated even under extreme occlusions. Our reconstruction method is markerless and can be applied to rigid as well as articulated objects. We introduce a new dataset consisting of close interactions between a human and an object and also evaluate on two scenes of humans performing martial arts. The experiments confirm the effectiveness of our framework and substantial improvements using 3D and novel view synthesis metrics compared to several existing approaches applicable in our setting.

Human activities are inherently complex, and even simple household tasks involve numerous object interactions. To better understand these activities and behaviors, it is crucial to model their dynamic interactions with the environment. The recent availability of affordable head-mounted cameras and egocentric data offers a more accessible and efficient means to understand dynamic human-object interactions in 3D environments. However, most existing methods for human activity modeling either focus on reconstructing 3D models of hand-object or human-scene interactions or on mapping 3D scenes, neglecting dynamic interactions with objects. The few existing solutions often require inputs from multiple sources, including multi-camera setups, depth-sensing cameras, or kinesthetic sensors. To this end, we introduce EgoGaussian, the first method capable of simultaneously reconstructing 3D scenes and dynamically tracking 3D object motion from RGB egocentric input alone. We leverage the uniquely discrete nature of Gaussian Splatting and segment dynamic interactions from the background. Our approach employs a clip-level online learning pipeline that leverages the dynamic nature of human activities, allowing us to reconstruct the temporal evolution of the scene in chronological order and track rigid object motion. Additionally, our method automatically segments object and background Gaussians, providing 3D representations for both static scenes and dynamic objects. EgoGaussian outperforms previous NeRF and Dynamic Gaussian methods in challenging in-the-wild videos and we also qualitatively demonstrate the high quality of the reconstructed models.

We introduce Joker, a new method for the conditional synthesis of 3D human heads with extreme expressions. Given a single reference image of a person, we synthesize a volumetric human head with the reference’s identity and a new expression. We offer control over the expression via a 3D morphable model (3DMM) and textual inputs. This multi-modal conditioning signal is essential since 3DMMs alone fail to define subtle emotional changes and extreme expressions, including those involving the mouth cavity and tongue articulation. Our method is built upon a 2D diffusion-based prior that generalizes well to out-of-domain samples, such as sculptures, heavy makeup, and paintings while achieving high levels of expressiveness. To improve view consistency, we propose a new 3D distillation technique that converts predictions of our 2D prior into a neural radiance field (NeRF). Both the 2D prior and our distillation technique produce state-of-the-art results, which are confirmed by our extensive evaluations. Also, to the best of our knowledge, our method is the first to achieve view-consistent extreme tongue articulation.

Current methods for extracting intrinsic image components, such as reflectance and shading, primarily rely on statistical priors. These methods focus mainly on simple synthetic scenes and isolated objects and struggle to perform well on challenging real-world data. To address this issue, we propose MLI-NeRF, which integrates Multiple Light information in Intrinsic-aware Neural Radiance Fields. By leveraging scene information provided by different light source positions complementing the multi-view information, we generate pseudo-label images for reflectance and shading to guide intrinsic image decomposition without the need for ground truth data. Our method introduces straightforward supervision for intrinsic component separation and ensures robustness across diverse scene types. We validate our approach on both synthetic and real-world datasets, outperforming existing state-of-the-art methods. Additionally, we demonstrate its applicability to various image editing tasks.

Scene flow, which provides the 3D motion field of the first frame from two consecutive point clouds, is vital for dynamic scene perception. However, contemporary scene flow methods face three major challenges. Firstly, they lack global flow embedding or only consider the context of individual point clouds before embedding, leading to embedded points struggling to perceive the consistent semantic relationship of another frame. To address this issue, we propose a novel approach called Dual Cross Attentive (DCA) for the latent fusion and alignment between two frames based on semantic contexts. This is then integrated into Global Fusion Flow Embedding (GF) to initialize flow embedding based on global correlations in both contextual and Euclidean spaces. Secondly, deformations exist in non-rigid objects after the warping layer, which distorts the spatiotemporal relation between the consecutive frames. For a more precise estimation of residual flow at next-level, the Spatial Temporal Re-embedding (STR) module is devised to update the point sequence features at current-level. Lastly, poor generalization is often observed due to the significant domain gap between synthetic and LiDAR-scanned datasets. We leverage novel domain adaptive losses to effectively bridge the gap of motion inference from synthetic to real-world. Experiments demonstrate that our approach achieves state-of-the-art (SOTA) performance across various datasets, with particularly outstanding results in real-world LiDAR-scanned situations.

Most recent 3D instance segmentation methods are open vocabulary, offering a greater flexibility than closed-vocabulary methods. Yet, they are limited to reasoning within a specific set of concepts, i.e. the vocabulary, prompted by the user at test time. In essence, these models cannot reason in an open-ended fashion, i.e., answering ``What are the objects in the scene?''. We introduce the first method to address 3D instance segmentation in a setting that is void of any vocabulary prior, namely a vocabulary-free setting. We leverage a large vision-language assistant and an open-vocabulary 2D instance segmenter to discover and ground semantic categories on the posed images. To form 3D instance mask, we first partition the input point cloud into dense superpoints, which are then merged into 3D instance masks. We propose a novel superpoint merging strategy via spectral clustering, accounting for both mask coherence and semantic coherence that are estimated from the 2D object instance masks. We evaluate our method using ScanNet200 and Replica, outperforming existing methods in both vocabulary-free and open-vocabulary settings. Code will be made available.

In this work, we address the challenge of accurate 3D human pose and shape (HPS) estimation from monocular images. The key to accuracy and robustness lies in high-quality training data. Existing training datasets containing real images with pseudo ground truth (pGT) use SMPLify to fit SMPL to sparse 2D joint locations, assuming a simplified camera with default intrinsics. We make two contributions that improve pGT accuracy. First, to estimate camera intrinsics, we develop a field-of-view prediction model HumanFoV trained on a dataset of images containing people. We use the estimated intrinsics to enhance the 4D-Humans dataset by incorporating a full perspective camera model during SMPLify fitting. Second, 2D joints provide limited constraints on 3D body shape, often resulting in average-looking bodies. To address this, we use the BEDLAM dataset to train a dense surface keypoint detector. We apply this detector to the 4D-Humans dataset and modify SMPLify to fit the detected keypoints, resulting in significantly more realistic body shapes. Finally, we enhance the HMR2.0 architecture to include the estimated camera parameters. We iterate the process of model training and SMPLify fitting initialized with the previously trained model. This leads to more accurate pGT and significant performance gains. Our method, CameraHMR, achieves state-of-the-art 3D accuracy on HPS benchmarks. Code will be available for research purposes.

This paper introduces GS-Pose, a unified framework for localizing and estimating the 6D pose of novel objects. GS-Pose begins with a set of posed RGB images of a previously unseen object and builds three distinct representations stored in a database. At inference, GS-Pose operates sequentially by locating the object in the input image, estimating its initial 6D pose using a retrieval approach, and refining the pose with a render-and-compare method. The key insight is the application of the appropriate object representation at each stage of the process. In particular, for the refinement step, we leverage 3D Gaussian splatting, a novel differentiable rendering technique that offers high rendering speed and relatively low optimization time. Off-the-shelf toolchains and commodity hardware, such as mobile phones, can be used to capture new objects to be added to the database. Extensive evaluations on the LINEMOD and OnePose-LowTexture datasets demonstrate excellent performance, establishing the new state-of-the-art. The source code is publicly available at https://github.com/dingdingcai/GSPose.

Creating personalized 3D animations with precise control and realistic head motions remains challenging for current speech-driven 3D facial animation methods. Editing these animations is especially complex and time consuming, requires precise control and typically handled by highly skilled animators. Most existing works focus on controlling style or emotion of the synthesized animation and cannot edit/regenerate parts of an input animation. They also overlook the fact that multiple plausible lip and head movements can match the same audio input. To address these challenges, we present 3DiFACE, a novel method for holistic speech-driven 3D facial animation. Our approach produces diverse plausible lip and head motions for a single audio input and allows for editing via keyframing and interpolation. Specifically, we propose a fully-convolutional diffusion model that can leverage the viseme-level diversity in our training corpus. Additionally, we employ a speaking-style personalization and a novel sparsely-guided motion diffusion to enable precise control and editing. Through quantitative and qualitative evaluations, we demonstrate that our method is capable of generating and editing diverse holistic 3D facial animations given a single audio input, with control between high fidelity and diversity.

Physically-based renderings contain Monte Carlo noise, with variance that increases as the number of rays per pixel decreases. This noise, while zero-mean for good modern renderers, can have heavy tails (most notably, for scenes containing specular or refractive objects). Learned methods for restoring low fidelity renders are highly developed, because suppressing render noise means one can save compute and use fast renders with few rays per pixel. We demonstrate that a diffusion model can denoise low fidelity renders successfully. Furthermore, our method can be conditioned on a variety of natural render information, and this conditioning helps performance. Quantitative experiments show that our method is competitive with SOTA across a range of sampling rates. Qualitative examination of the reconstructions suggests that the image prior applied by a diffusion method strongly favors reconstructions that are like real images -- so have straight shadow boundaries, curved specularities and no fireflies.

Although neural rendering has made significant advances in creating lifelike, animatable full-body and head avatars, incorporating detailed expressions into full-body avatars remains largely unexplored. We present DEGAS, the first 3D Gaussian Splatting (3DGS)-based modeling method for full-body avatars with rich facial expressions. Trained on multiview videos of a given subject, our method learns a conditional variational autoencoder that takes both the body motion and facial expression as driving signals to generate Gaussian maps in the UV layout. To drive the facial expressions, instead of the commonly used 3D Morphable Models (3DMMs) in 3D head avatars, we propose to adopt the expression latent space trained solely on 2D portrait images, bridging the gap between 2D talking faces and 3D avatars. Leveraging the rendering capability of 3DGS and the rich expressiveness of the expression latent space, the learned avatars can be reenacted to reproduce photorealistic rendering images with subtle and accurate facial expressions. Experiments on an existing dataset and our newly proposed dataset of full-body talking avatars demonstrate the efficacy of our method. We also propose an audio-driven extension of our method with the help of 2D talking faces, opening new possibilities for interactive AI agents.

The rise of Extended Reality (XR) requires efficient streaming of 3D online worlds, challenging current 3DGS representations to adapt to bandwidth-constrained environments. This paper proposes LapisGS, a layered 3DGS that supports adaptive streaming and progressive rendering. Our method constructs a layered structure for cumulative representation, incorporates dynamic opacity optimization to maintain visual fidelity, and utilizes occupancy maps to efficiently manage Gaussian splats. This proposed model offers a progressive representation supporting a continuous rendering quality adapted for bandwidth-aware streaming. Extensive experiments validate the effectiveness of our approach in balancing visual fidelity with the compactness of the model, with up to 50.71% improvement in SSIM, 286.53% improvement in LPIPS with 23% of the original model size, and shows its potential for bandwidth-adapted 3D streaming and rendering applications.

Point cloud registration approaches often fail when the overlap between point clouds is low due to noisy point correspondences. This work introduces a novel cross-attention mechanism tailored for Transformer-based architectures that tackles this problem, by fusing information from coordinates and features at the super-point level between point clouds. This formulation has remained unexplored primarily because it must guarantee rotation and translation invariance since point clouds reside in different and independent reference frames. We integrate the Gromov–Wasserstein distance into the cross-attention formulation to jointly compute distances between points across different point clouds and account for their geometric structure. By doing so, points from two distinct point clouds can attend to each other under arbitrary rigid transformations. At the point level, we also devise a self-attention mechanism that aggregates the local geometric structure information into point features for fine matching. Our formulation boosts the number of inlier correspondences, thereby yielding more precise registration results compared to state-of-the-art approaches. We have conducted an extensive evaluation on 3DMatch, 3DLoMatch, KITTI, and 3DCSR datasets. Project page: https://github.com/twowwj/FLAT.

Zero-shot 3D part segmentation is a challenging and fundamental task. In this work, we propose a novel pipeline, ZeroPS, which achieves high-quality knowledge transfer from 2D pretrained foundation models (FMs), SAM and GLIP, to 3D object point clouds. We aim to explore the natural relationship between multi-view correspondence and the FMs' prompt mechanism and build bridges on it. In ZeroPS, the relationship manifests as follows: 1) lifting 2D to 3D by leveraging co-viewed regions and SAM's prompt mechanism, 2) relating 1D classes to 3D parts by leveraging 2D-3D view projection and GLIP's prompt mechanism, and 3) enhancing prediction performance by leveraging multi-view observations. Extensive evaluations on the PartNetE and AKBSeg benchmarks demonstrate that ZeroPS significantly outperforms the SOTA method across zero-shot unlabeled and instance segmentation tasks. ZeroPS does not require additional training or fine-tuning for the FMs. ZeroPS applies to both simulated and real-world data. It is hardly affected by domain shift. The project page is available at https://luis2088.github.io/ZeroPS_page/.

This paper targets interactive object-level editing (\textit{e.g.}, deletion, recoloring, transformation, composition) in dynamic scenes. Recently, some methods aiming for flexible editing static scenes represented by neural radiance field (NeRF) have shown impressive synthesis quality, while similar capabilities in time-variant dynamic scenes remain limited. To solve this problem, we propose 4D-Editor, an interactive semantic-driven editing framework, allowing editing multiple objects in a dynamic NeRF with user strokes on a single frame. Specifically, we extend the original dynamic NeRF by incorporating Hybrid Semantic Feature Distillation to maintain spatial-temporal consistency after editing. In addition, a Recursive Selection Refinement module is presented to significantly boost object segmentation accuracy within a dynamic NeRF to aid the editing process. Moreover, we develop Multi-view Reprojection Inpainting to fill holes caused by incomplete scene capture after editing. Extensive quantitative and qualitative experiments on real application scenarios demonstrate that 4D-Editor achieves photo-realistic editing on dynamic NeRFs. Project page: \href{https://patrickddj.github.io/4D-Editor}{https://patrickddj.github.io/4D-Editor}

Matching a 2D contour to a non-rigidly deformed 3D mesh is a challenging problem due to ambiguities arising from dimensionality differences. In the past, product graph based methods were only able to either produce fast but noisy solutions, or smooth but slow solutions (the latter enabled by higher-order costs computed in the conjugate product graph). In this work, we propose an approximation of these higher-order costs so that they can be computed in the ordinary product graph. This leads to an efficient algorithm for high-quality 2D-3D shape matching and enables novel applications, like an interactive user interface which allows to refine the solution gradually. We show theoretically that our method is efficient, and we experimentally validate that the accuracy gap of our approximation to the optimum is small in practice. Our code is available.

We present a method for dynamic surface reconstruction of large-scale urban scenes from LiDAR. Depth-based reconstructions tend to focus on small-scale objects or large-scale SLAM reconstructions that treat moving objects as outliers. We take a holistic perspective and optimize a compositional model of a dynamic scene that decomposes the world into rigidly-moving objects and the background. To achieve this, we take inspiration from recent novel view synthesis methods and frame the reconstruction problem as a global optimization over neural surfaces, ego poses, and object poses, which minimizes the error between composed spacetime surfaces and input LiDAR scans. In contrast to view synthesis methods, which typically minimize 2D errors with gradient descent, we minimize a 3D point-to-surface error by coordinate descent, which we decompose into registration and surface reconstruction steps. Each step can be handled well by off-the-shelf methods without any re-training. We analyze the surface reconstruction step for rolling-shutter LiDARs, and show that deskewing operations common in continuous time SLAM can be applied to dynamic objects as well, improving results over prior art by 10X. Beyond pursuing dynamic reconstruction as a goal in and of itself, we propose that such a system can be used to auto-label partially annotated sequences and produce ground truth annotation for hard-to-label problems such as depth completion and scene flow.

Relative pose estimation for RGB-D cameras is crucial in a number of applications. A typical approach relies on RANSAC to find a triplet pair of 3D point correspondences from which relative pose can be derived. A key aspect to this work ensures the geometric consistency of the triplet, {\em i.e.}, pairwise distances between 3D points are preserved between the two views. Observe, however, that depth values are typically an order of magnitude less precise than feature locations, leading to large distance thresholds and admission of numerous false positives. This paper proposes that the constraint of 3D distance can be cast as a 2D constraint which we refer to as the {\em Geometric Correspondence Constraint (GCC)}. This constraint states that given one pair of correspondences, the two images are partitioned into a family nested curves such that corresponding points must lie on corresponding curves. This can act as a filter in the RANSAC process with significant savings in computation and with increased robustness and accuracy as demonstrated in experiments using TUM, ICL-NUIM, and RGBD Scene v2 datasets.