Reproduce “GASP: Unifying Geometric and Semantic Self-Supervised Pre-training for Autonomous Driving” 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
William Ljungbergh, Adam Lilja, Adam Tonderski. Arvid Laveno Ling, Carl Lindström, Willem Verbeke, Junsheng Fu, Christoffer Petersson, Lars Hammarstrand and 1 more
GASP introduces a self-supervised pre-training framework for autonomous driving, which unifies predictions of future geometry, semantics, and ego-path within a 4D environmental representation derived from unlabeled sensor data. The framework delivers enhanced performance on tasks such as semantic occupancy forecasting and online mapping, demonstrating an order of magnitude reduction in labeled data required for fine-tuning.
Key insights
- Unifying geometric, semantic, and ego-path self-supervised signals within a single continuous 4D representation allows for a more comprehensive and generalizable world model.
- Semantic supervision, specifically from distilled DINOv2 features, significantly enhances both geometric modeling accuracy and performance on downstream semantic-heavy tasks like map segmentation.
- The pre-trained GASP model enables strong performance in low-data fine-tuning regimes, requiring an order of magnitude less labeled data for tasks such as semantic BEV forecasting while maintaining competitive results.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.