Weiguang Zhao

Weiguang Zhao, Chaolong Yang, Jianan Ye et al. · nvidia

While weakly supervised multi-view face reconstruction (MVR) is garnering increased attention, one critical issue still remains open: how to effectively interact and fuse multiple image information to reconstruct high-precision 3D models. In this regard, we propose a novel pipeline called Deep Fusion MVR (DF-MVR) to explore the feature correspondences between multi-view images and reconstruct high-precision 3D faces. Specifically, we present a novel multi-view feature fusion backbone that utilizes face masks to align features from multiple encoders and integrates one multi-layer attention mechanism to enhance feature interaction and fusion, resulting in one unified facial representation. Additionally, we develop one concise face mask mechanism that facilitates multi-view feature fusion and facial reconstruction by identifying common areas and guiding the network's focus on critical facial features (e.g., eyes, brows, nose, and mouth). Experiments on Pixel-Face and Bosphorus datasets indicate the superiority of our model. Without 3D annotation, DF-MVR achieves 5.2% and 3.0% RMSE improvement over the existing weakly supervised MVRs respectively on Pixel-Face and Bosphorus dataset. Code will be available publicly at https://github.com/weiguangzhao/DF_MVR.

Weiguang Zhao, Yuyao Yan, Chaolong Yang et al.

Instance segmentation on point clouds is crucially important for 3D scene understanding. Most SOTAs adopt distance clustering, which is typically effective but does not perform well in segmenting adjacent objects with the same semantic label (especially when they share neighboring points). Due to the uneven distribution of offset points, these existing methods can hardly cluster all instance points. To this end, we design a novel divide-and-conquer strategy named PBNet that binarizes each point and clusters them separately to segment instances. Our binary clustering divides offset instance points into two categories: high and low density points (HPs vs. LPs). Adjacent objects can be clearly separated by removing LPs, and then be completed and refined by assigning LPs via a neighbor voting method. To suppress potential over-segmentation, we propose to construct local scenes with the weight mask for each instance. As a plug-in, the proposed binary clustering can replace traditional distance clustering and lead to consistent performance gains on many mainstream baselines. A series of experiments on ScanNetV2 and S3DIS datasets indicate the superiority of our model. In particular, PBNet ranks first on the ScanNetV2 official benchmark challenge, achieving the highest mAP. Code will be available publicly at https://github.com/weiguangzhao/PBNet.

Chaolong Yang, Yuyao Yan, Weiguang Zhao et al.

3D point clouds are rich in geometric structure information, while 2D images contain important and continuous texture information. Combining 2D information to achieve better 3D semantic segmentation has become mainstream in 3D scene understanding. Albeit the success, it still remains elusive how to fuse and process the cross-dimensional features from these two distinct spaces. Existing state-of-the-art usually exploit bidirectional projection methods to align the cross-dimensional features and realize both 2D & 3D semantic segmentation tasks. However, to enable bidirectional mapping, this framework often requires a symmetrical 2D-3D network structure, thus limiting the network's flexibility. Meanwhile, such dual-task settings may distract the network easily and lead to over-fitting in the 3D segmentation task. As limited by the network's inflexibility, fused features can only pass through a decoder network, which affects model performance due to insufficient depth. To alleviate these drawbacks, in this paper, we argue that despite its simplicity, projecting unidirectionally multi-view 2D deep semantic features into the 3D space aligned with 3D deep semantic features could lead to better feature fusion. On the one hand, the unidirectional projection enforces our model focused more on the core task, i.e., 3D segmentation; on the other hand, unlocking the bidirectional to unidirectional projection enables a deeper cross-domain semantic alignment and enjoys the flexibility to fuse better and complicated features from very different spaces. In joint 2D-3D approaches, our proposed method achieves superior performance on the ScanNetv2 benchmark for 3D semantic segmentation.

Weiguang Zhao, Haoran Xu, Xingyu Miao et al.

Point tracking aims to follow visual points through complex motion, occlusion, and viewpoint changes, and has advanced rapidly with modern foundation models. Yet progress toward general point tracking remains constrained by limited high-quality data, as existing datasets often provide insufficient diversity and imperfect trajectory annotations. To this end, we introduce SynthVerse, a large-scale, diverse synthetic dataset specifically designed for point tracking. SynthVerse includes several new domains and object types missing from existing synthetic datasets, such as animated-film-style content, embodied manipulation, scene navigation, and articulated objects. SynthVerse substantially expands dataset diversity by covering a broader range of object categories and providing high-quality dynamic motions and interactions, enabling more robust training and evaluation for general point tracking. In addition, we establish a highly diverse point tracking benchmark to systematically evaluate state-of-the-art methods under broader domain shifts. Extensive experiments and analyses demonstrate that training with SynthVerse yields consistent improvements in generalization and reveal limitations of existing trackers under diverse settings.

Weiguang Zhao, Guanyu Yang, Rui Zhang et al.

With the explosive 3D data growth, the urgency of utilizing zero-shot learning to facilitate data labeling becomes evident. Recently, methods transferring language or language-image pre-training models like Contrastive Language-Image Pre-training (CLIP) to 3D vision have made significant progress in the 3D zero-shot classification task. These methods primarily focus on 3D object classification with an aligned pose; such a setting is, however, rather restrictive, which overlooks the recognition of 3D objects with open poses typically encountered in real-world scenarios, such as an overturned chair or a lying teddy bear. To this end, we propose a more realistic and challenging scenario named open-pose 3D zero-shot classification, focusing on the recognition of 3D objects regardless of their orientation. First, we revisit the current research on 3D zero-shot classification, and propose two benchmark datasets specifically designed for the open-pose setting. We empirically validate many of the most popular methods in the proposed open-pose benchmark. Our investigations reveal that most current 3D zero-shot classification models suffer from poor performance, indicating a substantial exploration room towards the new direction. Furthermore, we study a concise pipeline with an iterative angle refinement mechanism that automatically optimizes one ideal angle to classify these open-pose 3D objects. In particular, to make validation more compelling and not just limited to existing CLIP-based methods, we also pioneer the exploration of knowledge transfer based on Diffusion models. While the proposed solutions can serve as a new benchmark for open-pose 3D zero-shot classification, we discuss the complexities and challenges of this scenario that remain for further research development. The code is available publicly at https://github.com/weiguangzhao/Diff-OP3D.

Weiguang Zhao, Rui Zhang, Qiufeng Wang et al.

3D semantic segmentation plays a fundamental and crucial role to understand 3D scenes. While contemporary state-of-the-art techniques predominantly concentrate on elevating the overall performance of 3D semantic segmentation based on general metrics (e.g. mIoU, mAcc, and oAcc), they unfortunately leave the exploration of challenging regions for segmentation mostly neglected. In this paper, we revisit 3D semantic segmentation through a more granular lens, shedding light on subtle complexities that are typically overshadowed by broader performance metrics. Concretely, we have delineated 3D semantic segmentation errors into four comprehensive categories as well as corresponding evaluation metrics tailored to each. Building upon this categorical framework, we introduce an innovative 3D semantic segmentation network called BFANet that incorporates detailed analysis of semantic boundary features. First, we design the boundary-semantic module to decouple point cloud features into semantic and boundary features, and fuse their query queue to enhance semantic features with attention. Second, we introduce a more concise and accelerated boundary pseudo-label calculation algorithm, which is 3.9 times faster than the state-of-the-art, offering compatibility with data augmentation and enabling efficient computation in training. Extensive experiments on benchmark data indicate the superiority of our BFANet model, confirming the significance of emphasizing the four uniquely designed metrics. Code is available at https://github.com/weiguangzhao/BFANet.

Xingyu Miao, Weiguang Zhao, Tao Lu et al.

Feed-forward multi-frame 3D reconstruction models often degrade on videos with object motion. Global-reference becomes ambiguous under multiple motions, while the local pointmap relies heavily on estimated relative poses and can drift, causing cross-frame misalignment and duplicated structures. We propose TrajVG, a reconstruction framework that makes cross-frame 3D correspondence an explicit prediction by estimating camera-coordinate 3D trajectories. We couple sparse trajectories, per-frame local point maps, and relative camera poses with geometric consistency objectives: (i) bidirectional trajectory-pointmap consistency with controlled gradient flow, and (ii) a pose consistency objective driven by static track anchors that suppresses gradients from dynamic regions. To scale training to in-the-wild videos where 3D trajectory labels are scarce, we reformulate the same coupling constraints into self-supervised objectives using only pseudo 2D tracks, enabling unified training with mixed supervision. Extensive experiments across 3D tracking, pose estimation, pointmap reconstruction, and video depth show that TrajVG surpasses the current feedforward performance baseline.

Weiguang Zhao, Xihao Guo, Tian Liang et al.

Robotic dexterous hands are central to contact-rich manipulation, with rapid progress driven by advances in hardware, sensing, control, simulation, and data generation. However, existing studies are often developed under different assumptions regarding hand embodiments, sensory configurations, task settings, training data, and evaluation protocols, making systematic comparison difficult and obscuring the developmental trajectory of the field. This survey provides a holistic review of dexterous hand research from four complementary aspects. First, we present a hardware-level analysis covering actuation, transmission, perception, and representative hand designs, highlighting the key trade-offs in force capability, compliance, bandwidth, integration, and system complexity. Furthermore, we review control and learning methods for dexterous manipulation from a methodological perspective, grouping representative works by major paradigms and tracing their evolution in chronological order. In addition, we consolidate datasets, modality design, and evaluation practices, which enables methodological progress to be interpreted together with the ways in which it is trained, benchmarked, and assessed. Finally, we discuss the major limitations of current dexterous hand research and summarize the corresponding future directions. By connecting hardware analysis, methodological development, data resources, and evaluation, this survey aims to provide a structured understanding of dexterous hand research and to clarify the most important open challenges for future study.

Chaolong Yang, Kai Yao, Yuyao Yan et al.

Audio-driven single-image talking portrait generation plays a crucial role in virtual reality, digital human creation, and filmmaking. Existing approaches are generally categorized into keypoint-based and image-based methods. Keypoint-based methods effectively preserve character identity but struggle to capture fine facial details due to the fixed points limitation of the 3D Morphable Model. Moreover, traditional generative networks face challenges in establishing causality between audio and keypoints on limited datasets, resulting in low pose diversity. In contrast, image-based approaches produce high-quality portraits with diverse details using the diffusion network but incur identity distortion and expensive computational costs. In this work, we propose KDTalker, the first framework to combine unsupervised implicit 3D keypoint with a spatiotemporal diffusion model. Leveraging unsupervised implicit 3D keypoints, KDTalker adapts facial information densities, allowing the diffusion process to model diverse head poses and capture fine facial details flexibly. The custom-designed spatiotemporal attention mechanism ensures accurate lip synchronization, producing temporally consistent, high-quality animations while enhancing computational efficiency. Experimental results demonstrate that KDTalker achieves state-of-the-art performance regarding lip synchronization accuracy, head pose diversity, and execution efficiency.Our codes are available at https://github.com/chaolongy/KDTalker.

Weiguang Zhao, Junting Dong, Rui Zhang et al.

Teleoperation is a key paradigm for transferring human dexterity to robots, yet most prior work targets objects that are initially static, such as grasping or manipulation. Dynamic object catch, where objects move before contact, remains underexplored. Pure teleoperation in this task often fails due to timing, pose, and force errors, highlighting the need for shared autonomy that combines human input with autonomous policies. To this end, we present Tele-Catch, a systematic framework for dexterous hand teleoperation in dynamic object catching. At its core, we design DAIM, a dynamics-aware adaptive integration mechanism that realizes shared autonomy by fusing glove-based teleoperation signals into the diffusion policy denoising process. It adaptively modulates control based on the interaction object state. To improve policy robustness, we introduce DP-U3R, which integrates unsupervised geometric representations from point cloud observations into diffusion policy learning, enabling geometry-aware decision making. Extensive experiments demonstrate that Tele-Catch significantly improves accuracy and robustness in dynamic catching tasks, while also exhibiting consistent gains across distinct dexterous hand embodiments and previously unseen object categories.

Weiguang Zhao, Chenru Jiang, Chengrui Zhang et al.

Given the diversity of devices and the product upgrades, cross-device research has become an urgent issue that needs to be tackled. To this end, we pioneer in probing the cross-device (cameras & robotics) grasping policy in the 3D open world. Specifically, we construct two real-world grasping setups, employing robotic arms and cameras from completely different manufacturers. To minimize domain differences in point clouds from diverse cameras, we adopt clustering methods to generate 3D object proposals. However, existing clustering methods are limited to closed-set scenarios, which confines the robotic graspable object categories and ossifies the deployment scenarios. To extend these methods to open-world settings, we introduce the SSGC-Seg module that enables category-agnostic 3D object detection. The proposed module transforms the original multi-class semantic information into binary semantic cues-foreground and background by analyzing the SoftMax value of each point, and then clusters the foreground points based on geometric information to form initial object proposals. Furthermore, ScoreNet‡ is designed to score each detection result, and the robotic arm prioritizes grasping the object with the highest confidence score. Experiments on two different types of setups highlight the effectiveness and robustness of our policy for cross-device robotics grasping research. Our code is provided in the supplementary and will be released upon acceptance.

Wuyang Luan, Junhui Li, Weiguang Zhao et al.

Vision-Language-Action (VLA) models based on flow matching -- such as pi0, pi0.5, and SmolVLA -- achieve state-of-the-art generalist robotic manipulation, yet their iterative denoising, typically 10 ODE steps, introduces substantial latency: on a modern GPU, denoising alone accounts for 80% of end-to-end inference time. Naively reducing the step count is unreliable, degrading success on most tasks due to the velocity field being uncalibrated for single-step jumps. We present SnapFlow, a plug-and-play self-distillation method that compresses multi-step denoising into a single forward pass (1-NFE) for flow-matching VLAs. SnapFlow mixes standard flow-matching samples with consistency samples whose targets are two-step Euler shortcut velocities computed from the model's own marginal velocity predictions, avoiding the trajectory drift caused by conditional velocities, as we analyze theoretically. A zero-initialized target-time embedding lets the network switch between local velocity estimation and global one-step generation within a single architecture. SnapFlow requires no external teacher, no architecture changes, and trains in ~12h on a single GPU. We validate on two VLA architectures spanning a 6x parameter range, with identical hyperparameters: on pi0.5 (3B) across four LIBERO suites (40 tasks, 400 episodes), SnapFlow achieves 98.75% average success -- matching the 10-step teacher at 97.75% and slightly exceeding it -- with 9.6x denoising speedup and end-to-end latency reduced from 274ms to 83ms; on SmolVLA (500M), it reduces MSE by 8.3% with 3.56x end-to-end acceleration. An action-step sweep on long-horizon tasks reveals that SnapFlow maintains its advantage across execution horizons, achieving 93% at n_act=5 where the baseline reaches only 90%. SnapFlow is orthogonal to layer-distillation and token-pruning approaches, enabling compositional speedups.

Jianan Ye, Weiguang Zhao, Xi Yang et al.

Point cloud anomaly detection under the anomaly-free setting poses significant challenges as it requires accurately capturing the features of 3D normal data to identify deviations indicative of anomalies. Current efforts focus on devising reconstruction tasks, such as acquiring normal data representations by restoring normal samples from altered, pseudo-anomalous counterparts. Our findings reveal that distributing attention equally across normal and pseudo-anomalous data tends to dilute the model's focus on anomalous deviations. The challenge is further compounded by the inherently disordered and sparse nature of 3D point cloud data. In response to those predicaments, we introduce an innovative approach that emphasizes learning point offsets, targeting more informative pseudo-abnormal points, thus fostering more effective distillation of normal data representations. We also have crafted an augmentation technique that is steered by normal vectors, facilitating the creation of credible pseudo anomalies that enhance the efficiency of the training process. Our comprehensive experimental evaluation on the Anomaly-ShapeNet and Real3D-AD datasets evidences that our proposed method outperforms existing state-of-the-art approaches, achieving an average enhancement of 9.0% and 1.4% in the AUC-ROC detection metric across these datasets, respectively.

Wuyang Luan, Junhui Li, Weiguang Zhao et al.

Visual navigation is a core challenge in Embodied AI, requiring autonomous agents to translate high-dimensional sensory observations into continuous, long-horizon action trajectories. While generative policies based on diffusion models and Schrödinger Bridges (SB) effectively capture multimodal action distributions, they require dozens of integration steps due to high-variance stochastic transport, posing a critical barrier for real-time robotic control. We propose Rectified Schrödinger Bridge Matching (RSBM), a framework that exploits a shared velocity-field structure between standard Schrödinger Bridges ($\varepsilon=1$, maximum-entropy transport) and deterministic Optimal Transport ($\varepsilon\to 0$, as in Conditional Flow Matching), controlled by a single entropic regularization parameter $\varepsilon$. We prove two key results: (1) the conditional velocity field's functional form is invariant across the entire $\varepsilon$-spectrum (Velocity Structure Invariance), enabling a single network to serve all regularization strengths; and (2) reducing $\varepsilon$ linearly decreases the conditional velocity variance, enabling more stable coarse-step ODE integration. Anchored to a learned conditional prior that shortens transport distance, RSBM operates at an intermediate $\varepsilon$ that balances multimodal coverage and path straightness. Empirically, while standard bridges require $\geq 10$ steps to converge, RSBM achieves over 94% cosine similarity and 92% success rate in merely 3 integration steps -- without distillation or multi-stage training -- substantially narrowing the gap between high-fidelity generative policies and the low-latency demands of Embodied AI.

Xinzhe Xia, Weiguang Zhao, Yuyao Yan et al.

3D open-world classification is a challenging yet essential task in dynamic and unstructured real-world scenarios, requiring both open-category and open-pose recognition. To address these challenges, recent wisdom often takes sophisticated 2D pre-trained models to provide enriched and stable representations. However, these methods largely rely on how 3D objects can be projected into 2D space, which is unfortunately not well solved, and thus significantly limits their performance. Unlike these present efforts, in this paper we make a pioneering exploration of 3D generative models for 3D open-world classification. Drawing on abundant prior knowledge from 3D generative models, we additionally craft a rotation-invariant feature extractor. This innovative synergy endows our pipeline with the advantages of being training-free, open-category, and pose-invariant, thus well suited to 3D open-world classification. Extensive experiments on benchmark datasets demonstrate the potential of generative models in 3D open-world classification, achieving state-of-the-art performance on ModelNet10 and McGill with 32.0% and 8.7% overall accuracy improvement, respectively.

Jingcheng Ni, Weiguang Zhao, Daniel Wang et al.

3D object removal is an important sub-task in 3D scene editing, with broad applications in scene understanding, augmented reality, and robotics. However, existing methods struggle to achieve a desirable balance among consistency, usability, and computational efficiency in multi-view settings. These limitations are primarily due to unintuitive user interaction in the source view, inefficient multi-view object mask generation, computationally expensive inpainting procedures, and a lack of applicability across different radiance field representations. To address these challenges, we propose a novel pipeline that improves the quality and efficiency of multi-view object mask generation and inpainting. Our method introduces an intuitive region-based interaction mechanism in the source view and eliminates the need for camera poses or extra model training. Our lightweight HoMM module is employed to achieve high-quality multi-view mask propagation with enhanced efficiency. In the inpainting stage, we further reduce computational costs by performing inpainting only on selected key views and propagating the results to other views via homography-based mapping. Our pipeline is compatible with a variety of radiance field frameworks, including NeRF and 3D Gaussian Splatting, demonstrating improved generalizability and practicality in real-world scenarios. Additionally, we present a new 3D multi-object removal dataset with greater object diversity and viewpoint variation than existing datasets. Experiments on public benchmarks and our proposed dataset show that our method achieves state-of-the-art performance while reducing runtime to one-fifth of that required by leading baselines.