Researchers have developed a new pipeline using the Segment Anything Model (SAM) to generate dense, pixel-level annotations for autonomous driving datasets that previously only had bounding boxes. This SAM-based approach has been applied to the Zenseact Open Dataset (ZOD), producing over 100,000 annotated frames, with a curated subset of 2,300 frames showing a 36% acceptance rate. Evaluations on transformer-based CLFT and CNN-based DeepLabV3+ architectures achieved up to 48.1% mIoU, with specialized models showing promise for rare classes. The pipeline was further validated on the Iseauto platform, reaching 77.5% mIoU, and demonstrated effective cross-sensor transfer learning. AI
IMPACT Enhances the quality and utility of datasets for autonomous driving research, potentially accelerating development of more robust perception systems.
RANK_REASON The cluster contains multiple arXiv papers detailing research into improving autonomous driving datasets and models.
- Adaptive Gated Location Refinement
- CULane
- CurveLanes
- LaneIoU-guided Confidence Calibration
- Iseauto
- Segment Anything Model
- Zenseact Open Dataset
AI-generated summary · Google Gemini · from 5 sources. How we write summaries →
