Video-Based Semi-automatic Drivable Area Segmentation

Video-Based Semi-automatic Drivable Area Segmentation

연구 분야: Verification

학회: International Conference on Pattern Recognition

Drivable area segmentation is a crucial task in autonomous driving. While current works mainly focus on analyzing single image and overlook the intra-video associations, due to the limited availability of video-based datasets. We present a novel prototype-based approach, named DASeg, to tackle the challenge of annotating intra-video query images using annotated support images as guidance. The primary obstacle lies in aggregating representative prototypes while ensuring resilience to variations in appearance and position across the video. Our method consists of three key components: position embedding for utilizing positional priors, soft-pooling for alleviating the limited coverage of intra-class variations from the support provided, and prototype regularization for generalizability enhancement. We augmented the lane detection dataset VIL-100 by incorporating drivable area annotations, resulting in a new dataset named VDA-100, which was employed to evaluate the performance of the proposed method. Experiments show that our method achieves mIoU score of 88.3% with the pre-trained backbone from lane detection model, and 89.1% when trained from scratch. Our code and dataset is available at https://github.com/CZY-Code/DASeg.