PartSLIP: Low-Shot Part Segmentation for 3D Point Clouds via Pretrained Image-Language Models

1University of California San Diego, 2Qualcomm AI Research
teaser image.

We propose PartSLIP, a zero/few-shot method for 3D point cloud part segmentation by leveraging pretrained image-language models. The figure shows our semantic segmentation results.

Abstract

Generalizable 3D part segmentation is important but challenging in vision and robotics. Training deep models via conventional supervised methods requires large-scale 3D datasets with fine-grained part annotations, which are costly to collect. This paper explores an alternative way for low-shot part segmentation of 3D point clouds by leveraging a pretrained image-language model, GLIP, which achieves superior performance on open-vocabulary 2D detection. We transfer the rich knowledge from 2D to 3D through GLIP-based part detection on point cloud rendering and a novel 2D-to-3D label lifting algorithm. We also utilize multi-view 3D priors and few-shot prompt tuning to boost performance significantly. Extensive evaluation on PartNet and PartNet-Mobility datasets shows that our method enables excellent zero-shot 3D part segmentation. Our few-shot version not only outperforms existing few-shot approaches by a large margin but also achieves highly competitive results compared to the fully supervised counterpart. Furthermore, we demonstrate that our method can be directly applied to iPhone-scanned point clouds without significant domain gaps.

Video

Overall Pipeline

The proposed components are highlighted in orange.

Instance Segmentation Results

PartSLIP (8-shot) results on the PartNet-Ensembled dataset. Different part instances are in different colors.

Real-World Point Clouds

PartSLIP can be directly applied to real-world point clouds without encountering significant domain gap: iPhone-scanned point clouds (first row), text prompts (second row), and PartSLIP (8-shot) results (third row).

Quantitative Comparison

PartSLIP achieves impressive zero-shot performances, and few-shot results are highly competitive compared to the fully supervised counterparts.

BibTeX

 @inproceedings{liu2023partslip,
        title={Partslip: Low-shot part segmentation for 3d point clouds via pretrained image-language models},
        author={Liu, Minghua and Zhu, Yinhao and Cai, Hong and Han, Shizhong and Ling, Zhan and Porikli, Fatih and Su, Hao},
        booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
        pages={21736--21746},
        year={2023}
      }