Object Detection in Foggy Scenes by Embedding Depth and Reconstruction into Domain Adaptation

Most existing domain adaptation (DA) methods align the features based on feature distributions and ignore aspects related to fog, background target objects, rendering suboptimal performance. In our DA framework, we retain depth information during alignment. A consistency loss between generated fog transmission map is introduced strengthen retention of in aligned features. To address false object potentially process, propose an encoder-decoder framework reconstruct fog-free image. This reconstruction also reinforces encoder, i.e., backbone, minimize Moreover, involve data training both module detection a semi-supervised manner, so that exposed unlabeled data, type used testing stage. Using these ideas, method significantly outperforms state-of-the-art (47.6 mAP against 44.3 Foggy Cityscapes dataset), obtains best performance multiple real-image public datasets. Code available at: https://github.com/VIML-CVDL/Object-Detection-in-Foggy-Scenes .