Lane delineation is essential information for autonomous vehicles. Being able to further differentiate between lane marking types gives the computational system a greater level of information regarding the road system and could improve safety assessment relying on infrastructure data and infrastructure management by itself. In this work, we propose a dual-branch weakly supervised pipeline for lane marking segmentation, relying on a Bootstrap Your Own Latent – BYOL self-supervised branch that uses contrastive loss to learn intrinsic patterns from the training dataset and a LinkNet segmentation branch for the lane marking segmentation itself; alternatively, BYOL self-supervision is used to fine-tune a pretrained segmentation model. The open dataset TuSimple is used as the base dataset for our analysis, and lane marking types are manually assigned after the labeled pixels are separated into individual entities based on their graph structure, forming an annotated ground truth dataset with noisy labels. Results show that fine-tuning with contrastive learning increases the semantic information learned by the segmentation model, leveraging more accurate lane delineation extraction, while weak supervision presents consistent learning but lower metric scoring. This work presents a novel discussion on the application of self-supervision for multiclass lane delineation with potential application in transport safety and management.
| ID | pc694 |
| Presentation | |
| Full Text | |
| Tags |
