Source-Free Unsupervised Domain Adaptation for Object Detection on Infrared Images

Object detection is an important task in various fields such as autonomous driving, robotics. However, when the model trained for object detection is deployed to another domain, performance drop can be observed. To deal with such issue, unsupervised domain adaptation for object detection aims to utilize unlabeled images in the target domain and labeled images in the source domain to prevent performance drop. Source-free unsupervised domain adaptation for object detection further points out the problem arising from the usage of source domain data, and utilizes only the source-trained model and unlabeled target domain data. However, a common approach in source-free unsupervised domain adaptation for object detection using pseudo-labels for self training has its weakness when facing a large domain difference, such as visible image to infrared image. This stems from the pseudo-labeling procedure, where only high confidence pseudo labels are utilized. In this paper, we propose a new algorithm to further utilize low confidence objects as pseudo-label, which makes the self-training capable of boosting the performance on large domain gaps. We validate the effectiveness of our algorithm by quantitatively and qualitatively comparing our algorithm with other methods on a visible image to infrared image domain shift scenario.