Accurate lane delineation is essential for road safety assessment, infrastructure management, and autonomous driving applications. In this work, we propose a weakly supervised lane segmentation pipeline to eliminate the need for costly manual annotations by integrating noisy label generation with self-supervised learning. Initial noisy labels are generated using a rule-based computer vision approach based on color and shape cues from street-level RGB imagery. These labels supervise a LinkNet-based segmentation model, while a parallel BYOL (Bootstrap Your Own Latent) self-supervised branch refines latent representations through contrastive learning. The dual-branch model is trained end-to-end by jointly optimizing segmentation and self-supervised objectives, optimizing the network by expanding the features from noisy labels. Experiments were conducted on a private dataset of road images collected in Northern Italy and on the public dataset TUSimple, which was also used to validate our pipeline using the annotated labels. Initial results after training both the weakly supervised model and the fully supervised baseline for 20 epochs showed that, while the baseline stabilized earlier, the proposed approach shows continued learning and captured the underlying components of lane markings rather than reproduced the labeled output. To understand better the impact of adding the self-supervised branch, we compared both models under different conditions: establishing a no-improvement threshold to consider the model fully trained and using the weak supervision for fine tuning instead of full training for 20, 50, 100 and 300 epochs. Results showed that fine tuning with contrastive learning avoided over-fitting and consistently presented better transferability.
| ID | pc667 |
| Presentation | |
| Tags |
