Domain Adaptive Detector via Variational Inference
Hwa-Rang Kim, Kwang‐Ju Kim, Doo-Hyun Choi · 2022
Recent deep learning-based models work enough for object detection. However, there are still challenging tasks aiming to address domain shift problem caused when a model is applied to target data that have different domains from source data. In this paper, we are concerned with a scenario in domain generalization to aim to perform well on unseen target domains. We propose a novel single domain generalization method using variational inference to improve cross-domain robustness for object detection. We build a model that has latent features following a prior distribution to achieve feature alignment. Specifically, the latent features are guided to follow Gaussian distribution for arbitrary inputs, hence the model can be domain-invariant. We utilize Faster R-CNN as a base detector, and our proposed approach is evaluated on Cityscapes, KITTI, SIM10k datasets. The results show that our method can provides a general solution for cross-domain robustness, only using source data in tackling single domain generalization.