Reproduce “Faster Training, Fewer Labels: Self-Supervised Pretraining for Fine-Grained BEV Segmentation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Daniel Busch, Christian Bohn, Thomas Kurbiel, Klaus Friedrichs, Richard Meyes, Tobias Meisen

Researchers at the University of Wuppertal and APTIV developed a two-phase training strategy for fine-grained Bird's Eye View (BEV) semantic segmentation that reduces reliance on expensive BEV ground truth annotations. The method achieves up to a 2.5 percentage point increase in mIoU over fully supervised baselines while using only 50% of BEV labels and reducing total training time by up to two-thirds.

Key insights

  • Self-supervised pretraining effectively establishes robust geometric projection and feature lifting capabilities from camera images to BEV, even when pretraining only qualitatively matches pseudo-labels, leading to rapid performance gains during fine-tuning.
  • Reducing BEV ground truth annotations by 50% during fine-tuning, when combined with self-supervised pretraining, can lead to superior performance compared to fully supervised baselines.
  • The temporal loss, while offering modest improvements during pretraining, demonstrates its primary value in mitigating occlusions and blind spots, effectively acting as a regularization in the initial learning phase.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.