19 July 2024 in ACM Transactions on Graphics
We present BlockFusion, a diffusion-based model that generates 3D scenes as unit blocks and seamlessly incorporates new blocks to extend the scene. BlockFusion is trained using datasets of 3D blocks that are randomly cropped from complete 3D scene meshes. Through per-block fitting, all training blocks are converted into the hybrid neural fields: with a tri-plane containing the geometry features, followed by a Multi-layer Perceptron (MLP) for decoding the signed distance values. A variational auto-encoder is employed to compress the tri-planes into the latent tri-plane space, on which the denoising diffusion process is performed. Diffusion applied to the latent representations allows for high-quality and diverse 3D scene generation.To expand a scene during generation, one needs only to append empty blocks to overlap with the current scene and extrapolate existing latent tri-planes to populate new blocks. The extrapolation is done by conditioning the generation process with the feature samples from the overlapping tri-planes during the denoising iterations. Latent tri-plane extrapolation produces semantically and geometrically meaningful transitions that harmoniously blend with the existing scene. A 2D layout conditioning mechanism is used to control the placement and arrangement of scene elements. Experimental results indicate that BlockFusion is capable of generating diverse, geometrically consistent and unbounded large 3D scenes with unprecedented high-quality shapes in both indoor and outdoor scenarios.
14 November 2019 in IEEE Transactions on Intelligent Transportation Systems
Traditional mapping modules in Visual Simultaneous Localization and Mapping (i.e. Visual SLAM) systems can only estimate 3D information of isolated sparse or semi-dense feature points. But there are lots of object instances in the environments which geometric information can be utilized to enhance the quality of mapping and localization. Hence, it is required for the Visual SLAM system to utilize high-dimensional features like object instances or structural lines in mapping and localization. To meet the gap between the above requirements and the traditional implementation of Visual SLAM systems, we present in this paper a novel Visual SLAM method that can effectively utilize texture-less object instances for mapping and localization. The proposed Visual SLAM method includes newly designed feature extraction, matching, localization and mapping modules, which jointly use object features and point features to estimate camera 6-DOF poses and do richer map construction. A group of organized raster points is used to represent objects during feature matching and pose estimation process in the proposed Visual SLAM pipeline. Owing to the object feature fusion in the co-visibility graph it could conduct scale aware bundle adjustments to reduce accumulated error. The advantages of proposed Visual SLAM method are demonstrated through experiments conducted both on synthetic datasets and real-world datasets.
24 July 2020 in IEEE Transactions on Circuits and Systems for Video Technology
A reliable and accurate visual object 6DoF pose tracking system for texture-less objects plays an important role in various fields of modern industry hence it has been a hot research topic for decades. Traditional feature-based pose tracking methods require rich feature points on objects, and it cannot handle texture-less objects. To tackle this problem, the paper proposes a novel edge-based method for continuous 6DOF pose tracking of texture-less objects. The pose of the object is estimated by minimizing the matching error between extracted image edges and re-projected CAD model edges. The matching error is represented using the Directional Chamfer Matching (DCM) Tensor. Compared with previous methods, the proposed system improves the overall performance of pose tracking system in two ways. Firstly, the method proposes an analytical mathematic model in the optimization process; Secondly, the method proposes an adaptive edge point weighting algorithm to tackle the occlusion or edge weakness problem. Both methods help improve the accuracy and robustness of the pose estimation system. With the benefit of GPU acceleration on DCM Tensor calculation the proposed method could run in real time. Extensive experiments are conducted both on public datasets and CG-rendered synthetic datasets to validate the accuracy and robustness of the proposed algorithm against other state-of-the-art methods.
16 August 2022 in IEEE Transactions on Instrumentation and Measurement
Visual simultaneous localization and mapping (VSLAM) is an appropriate method for positioning and navigation of intelligent unmanned systems under Global Navigation Satellite Systems (GNSS)-denied environment, but it is still facing some dilemmas in repetitive large-scale environments. In this article, a VSLAM method based on bifocal-binocular vision is proposed. By introducing the binocular camera with different focal lengths, the perception ability of the system in vast space is improved as the designed cameras could complement each other at different working distances. Meanwhile, considering the inherent structure of the scene, additional optimization is proposed to reduce the accumulated error based on the markers distribution knowledge obtained from online placement inference. The algorithm proposed in this article significantly improves the stability and accuracy of the VSLAM system in repetitive large-scale scenes, and is validated in both virtual datasets and real-world environments.
6 January 2023 in IEEE Robotics and Automation Letters
Monocular depth estimation is an ambiguous problem, thus global structural cues play an important role in current data-driven single-view depth estimation methods. Panorama images capture the complete spatial information of their surroundings utilizing the equirectangular projection which introduces large distortion. This requires the depth estimation method to be able to handle the distortion and extract global context information from the image. In this paper, we propose an end-to-end deep network for monocular panorama depth estimation on a unit spherical surface. Specifically, we project the feature maps extracted from equirectangular images onto unit spherical surface sampled by uniformly distributed grids, where the decoder network can aggregate the information from the distortion-reduced feature maps. Meanwhile, we propose a global cross-attention-based fusion module to fuse the feature maps from skip connection and enhance the ability to obtain global context. Experiments are conducted on five panorama depth estimation datasets, and the results demonstrate that the proposed method substantially outperforms previous state-of-the-art methods. All related codes will be open-sourced in the upcoming days.
1 July 2026 in IEEE Transactions on Visualization and Computer Graphics
While recent advancements have demonstrated remarkable progress in general 3D shape generation, the challenge of automatically generating wearable 3D assets remains largely unexplored. To address this gap, we present BAG - a Body-aligned Asset Generation method that produces 3D wearable assets which can be automatically fitted onto given 3D human bodies. This is achieved by controlling the 3D generation process using human body shape and pose information. Specifically, we first construct a general single-image-to-consistent-multi-view diffusion model, and train it on the large-scale Objaverse dataset to ensure diversity and generalizability. We then train a body-conditioned multi-view ControlNet to guide the generator toward producing body-aligned multi-view images. The control signal leverages multi-view 2D projections of the target human body, where pixel values represent the XYZ coordinates of the body surface in a canonical space. The resulting body-conditioned multi-view diffusion outputs body-aligned images, which are subsequently fed into a native 3D diffusion model to reconstruct the 3D shape of the asset. Finally, we recover the similarity transformation using multi-view silhouette supervision and mitigate asset-body penetration using physics-based simulation, ensuring accurate asset fitting onto the target body. Experimental results demonstrate that our method significantly outperforms existing approaches in terms of prompt adherence, shape diversity, and shape quality.