Source-Free Domain Adaptation for Point Cloud Semantic Segmentation

Jianshe Duan, Yachao Zhang, Yanyun Qu

Abstract

Point cloud semantic segmentation (PCSS) is fundamental in 3D scene understanding. Domain adaptation methods for PCSS enable transferring the knowledge learned from a labeled source domain to an unlabeled target domain with different data distributions. However, they become inapplicable in privacy-preserving scenarios, where we only can leverage a given source model instead of the source data. In this paper, we propose the first source-free domain adaptation PCSS framework to make a given source model generalize to the target domain well. We found existing self-training based source-free domain adaptation methods inevitably lead to confirmation bias and suffer from serious model degradation on PCSS. Therefore, we devise a novel Teacher-Guide Source-Free (TGSF) framework to conquer the above two challenges, including the bidirectional pseudo-label selection and multi-level target consistency learning. Extensive experiments on three benchmarks verify that TGSF achieves state-of-the-art performance, and even outperforms those methods that access the source data.
BibTeX
@inproceedings{duan2024source,
  author = {Jianshe Duan and Yachao Zhang and Yanyun Qu},
  title = {{Source-Free Domain Adaptation for Point Cloud Semantic Segmentation}},
  booktitle = {ICME},
  year = {2024},
  doi = {10.1109/icme57554.2024.10687489},
}

Loading related content...

Figure