PanopticRecon: Leverage Open-vocabulary Instance Segmentation for Zero-shot Panoptic Reconstruction

IROS 2024 (Oral)

Xuan Yu, Yili Liu, Chenrui Han, Sitong Mao, Shunbo Zhou, Rong Xiong, Yiyi Liao, Yue Wang*
Zhejiang University, Huawei Cloud Computing Technologies Co., Ltd.
*Corresponding author
Paper Video Code

Abstract

Panoptic reconstruction is a challenging task in 3D scene understanding. However, most existing methods heavily rely on pre-trained semantic segmentation models and known 3D object bounding boxes for 3D panoptic segmentation, which is not available for in-the-wild scenes. In this paper, we propose a novel zero-shot panoptic reconstruction method from RGB-D images of scenes. For zero-shot segmentation, we leverage open-vocabulary instance segmentation, but it has to face partial labeling and instance association challenges. We tackle both challenges by propagating partial labels with the aid of dense generalized features and building a 3D instance graph for associating 2D instance IDs. Specifically, we exploit partial labels to learn a classifier for generalized semantic features to provide complete labels for scenes with dense distilled features. Moreover, we formulate instance association as a 3D instance graph segmentation problem, allowing us to fully utilize the scene geometry prior and all 2D instance masks to infer global unique pseudo 3D instance ID. Our method outperforms state-of-the-art methods on the indoor dataset ScanNet V2 and the outdoor dataset KITTI-360, demonstrating the effectiveness of our graph segmentation method and reconstruction network.

Approach Overview

Teaser PanopticRecon consists of a reconstruction task and a segmentation task. The first step of the reconstruction task realizes the implicit surface reconstruction through RGB-D observations to provide the scene geometry for the segmentation task. Secondly, the segmentation task builds a graph from the normal of mesh, and infers 3D pseudo IDs to associate the 2D instance IDs by instance mask of Grounded SAM. In addition, 3D instance ID corrects some of the erroneous semantic labels. Then, the second reconstruction step realizes 2D-3D labeling supervised by consistent semantic and instance labels, and finally obtains the panoptic mesh, point cloud, and novel view images of the scene.

Teaser

Graph Inference

The points in the upper graph in \textbf{(a)} are the nodes (superpoints) of the graph. The color region corresponding to each node is the superface. We determine the nodes in an instance mask of a frame based on the overlap between the instance mask provided by Grounded SAM and the mask projected by the superface, and vote for the edges between the selected nodes. Similarly, we reduce the votes for the edges between nodes corresponding to masks of different instances in that frame. The edges with non-positive votes are finally cut and the nodes connected form an instance as shown in \textbf{(b)}. Once the 3D instance pseudo IDs are obtained, we associate 2D instance IDs while correcting incorrect semantic labels.
Teaser

Network

Based on the reconstruction architecture of a multi-level hash representation grid and small MLP, multi-branch joint optimization achieves multiple tasks of SDF reconstruction, appearance reconstruction, feature distillation, semantic segmentation, and instance segmentation. The loss function of the optimization process is not only the basic loss of SDF and color, but we also set the DINO-v2 feature vector L2-loss to unify the mesh features of similar objects or backgrounds and learn small classifiers targeting the mesh features through the cross-entropy loss of semantics and instances, to achieve the unsupervised situation of some regions due to blank labels to a certain extent.

BibTeX


        @article{yu2024panopticrecon,
          title={PanopticRecon: Leverage Open-vocabulary Instance Segmentation for Zero-shot Panoptic Reconstruction},
          author={Yu, Xuan and Liu, Yili and Han, Chenrui and Mao, Sitong and Zhou, Shunbo and Xiong, Rong and Liao, Yiyi and Wang, Yue},
          journal={arXiv preprint arXiv:2407.01349},
          year={2024}
          }