CLSD: Continual Learning for Lane Line Segmentation Across Domains

Lane line segmentation is a fundamental task in the field of autonomous driving, and it’s intuitive that automatic vehicles should have the ability of processing data from different domains. The usual assumption is that training data of all the domains is available at a time, however, it’s more common to collect data continually and train the visual system incrementally. To date, many works focus on segmenting lane line in a single domain, but no work considers continually segmenting lane line across domains. To tackle with this task, we presents a method, named as CLSD
¹
which leverages data from new domain while avoiding catastrophic forgetting. The core of CLSD are a training framework that effectively transfer knowledge from the previous model to the new model and a frustrating simple yet useful self-training strategy. We conduct extensive experiments based on TSD datasets and showcase that on daytime our training framework increases the miou by 10% on the fog domain data, and the self-straining increases the miou by 11% on the fog domain data and 47% in night domain data.
¹
Continual Learning for Lane Line Segmentation Across Domains