Reproduce “Semi-Supervised Diversity-Aware Domain Adaptation for 3D Object detection” on OpenResearch

About this paper

Bartłomiej Olber, Jakub Winter, Paweł Wawrzyński, Andrii Gamalii, Daniel Górniak, Marcin Łojek, Robert Nowak, Krystian Radlak

The method enables 3D LiDAR object detectors to adapt to new operational design domains using an exceptionally small amount of labeled target data, like 10 frames, achieving performance competitive with or superior to state-of-the-art methods that use significantly larger datasets. It incorporates a diversity-aware sample selection algorithm and post-training strategies to efficiently bridge domain gaps.

Key insights

• State-of-the-art 3D object detection domain adaptation can be achieved with an extremely limited annotation budget from the target domain, significantly reducing practical deployment costs.
• The proposed diversity-aware sample selection algorithm based on neuron activation patterns is crucial for maximizing learning efficiency from minimal target data, outperforming random selection.
• Post-training strategies such as L2-SP regularization and linear fading learning rates are more effective than standard fine-tuning or linear probing in adapting models and mitigating catastrophic forgetting.

Summary by alphaXiv.