Overview
BPLane strengthens three stages that directly shape final lane-detection performance: proposal parameterization, global-to-dense feature extraction, and output-cardinality selection after NMS.
Project Page
BPLane
BPLane is a proposal-refinement lane detector built around three coordinated improvements: a DINOv3+UPLiFT global-to-dense feature path, BiAngle proposal geometry, and a token-based lane-count prior with MAP-annealed decoding.
This project page is the home for the ETRI-Lane dataset and benchmark, together with BPLane pretrained weights, training code, and project updates.
ETRI-Lane contains lane data collected from South Korean cities including Daejeon, Seoul, Hwaseong, Ulsan, and Sejong, covering snow, rain, night driving, construction zones, and unpaved roads.
Project Resources
- ETRI-Lane dataset and benchmark resources
- BPLane pretrained weights
- BPLane training code and related updates