Multi-Domain Conditional Image Translation: Translating Driving Datasets From Clear-Weather to Adverse Conditions

Vishal Vinod, K. Ram Prabhakar, R. Venkatesh Babu, Anirban Chakraborty; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2021, pp. 1571-1582

Vision systems for fully autonomous navigation must perform well even in unstructured and degraded scenarios. In most driving datasets today, there is a bias toward clear-weather conditions as compared with extreme-weather owing to the difficulty in capturing and annotating large-scale image datasets degraded by adverse weather. While there has been extensive research on techniques such as deraining, dehazing and on tasks such as segmentation and domain adaptation, there has been minimal attention toward methods to effectively translate clear-weather driving datasets to extreme-weather domains. To address this, we present a method that builds on recent advances in Generative Networks and Self-Supervised Learning to perform conditional multi-domain image translation. We evaluate our method on the semantic scene understanding task and demonstrate quantitatively superior translation results from clear-weather conditions to adverse-weather shifted domains such as Rain, Night and Fog conditions. From our experiments, we show improved domain invariant content disentanglement, and segmentation methods trained with datasets translated using the proposed method have improved performance over single and multi-domain image translation baselines on real-world adverse weather data.