Train Till You Drop: Towards Stable and Robust Source-free Unsupervised 3D Domain Adaptation

We tackle the challenging problem of source-free unsupervised domainadaptation (SFUDA) for 3D semantic segmentation. It amounts to performingdomain adaptation on an unlabeled target domain without any access to sourcedata; the available information is a model trained to achieve good performanceon the source domain. A common issue with existing SFUDA approaches is thatperformance degrades after some training time, which is a by product of anunder-constrained and ill-posed problem. We discuss two strategies to alleviatethis issue. First, we propose a sensible way to regularize the learningproblem. Second, we introduce a novel criterion based on agreement with areference model. It is used (1) to stop the training when appropriate and (2)as validator to select hyperparameters without any knowledge on the targetdomain. Our contributions are easy to implement and readily amenable for allSFUDA methods, ensuring stable improvements over all baselines. We validate ourfindings on various 3D lidar settings, achieving state-of-the-art performance.The project repository (with code) is: github.com/valeoai/TTYD.