Multi-Model Learning for Real-Time Automotive Semantic Foggy Scene Understanding via Domain Adaptation

Robust semantic scene segmentation for automotive applications is a\nchallenging problem in two key aspects: (1) labelling every individual scene\npixel and (2) performing this task under unstable weather and illumination\nchanges (e.g., foggy weather), which results in poor outdoor scene visibility.\nSuch visibility limitations lead to non-optimal performance of generalised deep\nconvolutional neural network-based semantic scene segmentation. In this paper,\nwe propose an efficient end-to-end automotive semantic scene understanding\napproach that is robust to foggy weather conditions. As an end-to-end pipeline,\nour proposed approach provides: (1) the transformation of imagery from foggy to\nclear weather conditions using a domain transfer approach (correcting for poor\nvisibility) and (2) semantically segmenting the scene using a competitive\nencoder-decoder architecture with low computational complexity (enabling\nreal-time performance). Our approach incorporates RGB colour, depth and\nluminance images via distinct encoders with dense connectivity and features\nfusion to effectively exploit information from different inputs, which\ncontributes to an optimal feature representation within the overall model.\nUsing this architectural formulation with dense skip connections, our model\nachieves comparable performance to contemporary approaches at a fraction of the\noverall model complexity.\n