Towards Weakly-Supervised Domain Adaptation for Lane Detection
Authors:
- 17 June 2024
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 3553-3563
Abstract
Lane detection plays an indispensable role in automated driving functions and advanced driver assistance systems by providing fundamental spatial orientation, which is imperative for trajectory planning with traffic regulation compliance. The variability of lane structures across the world poses challenges for data-driven lane detection models. However, acquiring vast amounts of labeled data encompassing a wide variety of real-world scenarios for supervised learning is often cost-prohibitive. In this work, we propose a Weakly Supervised Domain Adaptation framework for Lane Detection (WSDAL), which requires easily-provided labels exclusively for the number of lanes in the target domain to aid the adaptation process. WSDAL consists of a teacher-student network, an additional segmentation head as an auxiliary task during training, and a novel loss function that incorporates the number of lanes prediction. As a versatile framework, WSDAL can be applied to any anchor-based lane detector. Between three frequently-used lane detection datasets (TuSimple, CULane and CurveLanes) for domain adaptation, WSDAL framework demonstrates its effectiveness and efficiency over common unsupervised domain adaptation methods and fully supervised training. In addition, we discuss the quality requisites from the labels for the weakly-supervised domain adaptation, indicating that label errors at realistic scales still provide satisfactory results on the considered tasks.Keywords
This publication has 54 references indexed in Scilit:
- A review of lane detection methods based on deep learningPattern Recognition, 2021
- Synthetic-to-Real Domain Adaptation for Lane Detection2021
- Ultra Fast Structure-Aware Deep Lane Detection2020
- Two-Phase Pseudo Label Densification for Self-training Based Domain Adaptation2020
- CurveLane-NAS: Unifying Lane-Sensitive Architecture Search and Adaptive Point Blending2020
- C-WSL: Count-Guided Weakly Supervised Localization2018
- Adaptive Batch Normalization for practical domain adaptationPattern Recognition, 2018
- ScribbleSup: Scribble-Supervised Convolutional Networks for Semantic Segmentation2016
- Learning Deep Features for Discriminative Localization2016
- Distinctive Image Features from Scale-Invariant KeypointsInternational Journal of Computer Vision, 2004