SAM-Enhanced Segmentation on Road Datasets: Balancing Critical Classes in Autonomous Driving
Researchers have developed a new method to create dense, pixel-level annotations for autonomous driving datasets that previously only had bounding boxes. This pipeline utilizes the Segment Anything Model (SAM) to convert bounding boxes into semantic masks, significantly enhancing the usability of datasets like the Zenseact Open Dataset (ZOD). The annotated data was used to evaluate transformer-based and CNN-based architectures, achieving up to 48.1% mIoU, and specialized models were explored to address extreme class imbalance for rare but critical objects like pedestrians and signs. AI
IMPACT Enables more robust training and evaluation of perception models for autonomous driving by providing high-quality, dense annotations.