# Uncertainty-aware Mean Teacher for Source-free Unsupervised Domain Adaptive 3D Object Detection

**Authors:** Deepti Hegde, Vishwanath Sindagi, Velat Kilic, A. Brinton Cooper, Mark Foster, Vishal Patel  
**arXiv:** [2109.14651](https://arxiv.org/abs/2109.14651)  
**Date:** 2021-09-29  
**Summary page:** https://www.summarizepaper.com/en/paper/2109.14651v1/

## Key points

- Introduces a novel approach to improve pseudo-label based self-training methods in unsupervised domain adaptation
- Addresses the issue of incorrect pseudo-labels generated by the source trained model
- Proposes an uncertainty-aware mean teacher framework to filter out incorrect pseudo-labels during training
- Down-weighs losses corresponding to uncertain pseudo-labels to avoid reinforcing erroneous predictions
- Incorporates automatic soft-sampling of pseudo-labeled data for improved accuracy and performance
- Demonstrated effectiveness through experiments on various domain adaptation scenarios
- Achieves state-of-the-art performance on the KITTI lidar target dataset

## Summary

The paper titled "Uncertainty-aware Mean Teacher for Source-free Unsupervised Domain Adaptive 3D Object Detection" introduces a novel approach to improve the efficacy of pseudo-label based self-training methods in source-free unsupervised domain adaptation. These methods heavily rely on the quality of labels generated by the source trained model, which can be incorrect with high confidence, making thresholding methods ineffective. To address this issue and prevent the reinforcement of errors caused by label noise, the authors propose an uncertainty-aware mean teacher framework. This framework leverages model uncertainty to implicitly filter out incorrect pseudo-labels during training. By down-weighing losses corresponding to uncertain pseudo-labels, the mean teacher network performs implicit filtering and avoids reinforcing erroneous predictions. The proposed method also incorporates automatic soft-sampling of pseudo-labeled data while aligning predictions from both the student and teacher networks. This soft-sampling process further improves the accuracy of pseudo-labels and enhances the overall performance of the domain adaptation system. The effectiveness of the uncertainty-aware mean teacher framework is demonstrated through experiments on various domain adaptation scenarios, including cross-dataset and cross weather conditions. The results show that this approach achieves state-of-the art performance on the KITTI lidar target dataset. In summary, this paper presents a novel technique that addresses label noise in pseudo label based self training approaches for unsupervised domain adaptation. By leveraging model uncertainty and performing implicit filtering, the proposed method significantly improves the quality of pseudo labels and achieves impressive results in various challenging domain adaptation scenarios.