Abstract:
Lane detection approaches relying on abundant annotations have achieved great progress in autonomous driving. These approaches may suffer from performance collapse when d...Show MoreMetadata
Abstract:
Lane detection approaches relying on abundant annotations have achieved great progress in autonomous driving. These approaches may suffer from performance collapse when directly applied to unseen images due to domain discrepancy. Most researches utilize unsupervised domain adaptation (UDA) to mitigate domain shifts. Despite this, there is still a large performance gap between UDA and supervised methods because lack of target supervision makes improving performance of cross-domain lane detection difficult. In this paper, we introduce semi-supervised domain adaptation (SSDA) to lane detection and develop a simple yet effective dual-adversarial learning scheme to minimize domain discrepancy. Specifically, we conduct global alignment by calculating entropy information of unlabeled target and source data to align entropy distribution and enhance feature confidence of lanes. Domain consistency adversarial learning is proposed to make full use of both source and unlabeled/labeled target data. A triplet of features is used to learn lane-invariant representations by distinguishing whether the feature belongs to same domain. This is conducive to promote consistency of lanes information in cross domain. In two datasets, our method obtains a better performance of 42.5% (54.9%) and 45.8% (64.1%) based on a small amount of target supervision. Experimental results demonstrate effectiveness of our method as well as its superiority to other methods in terms of reducing performance gap with supervised methods (https://github.com/shenhuqiji/LD-SSDA) Note to Practitioners—This paper focuses on the cross domain lane detection task to narrow domain discrepancy in lane detection among different scenarios for autonomous driving. It aims to reduce extra label annotations and retraining costs. Considering unsupervised domain adaptation method is hard to further improve detection performance when a large domain shift occurs in adaptation scenarios, we construct an effective dual-adversarial l...
Published in: IEEE Transactions on Automation Science and Engineering ( Early Access )