Unsupervised Domain Adaptive Object Detection Based on Channel Recalibration and Ancillary Net

Abstract

In the field of object detection, it is generally hoped that the model trained in the label-rich environment can be well applied to other various and label-agnostic environment. However, due to the existence of domain discrepancy between the training environment and the application environment, the performance of conventional object detection models inevitably encounters a sharp drop. Aiming at the limitations of the object detection application environment and the time-consuming of image labeling, this paper proposes an unsupervised domain adaptive object detection model based on Faster R-CNN. We integrate SENets into the feature extraction network to capture channel-wise correlations between features and strengthen the representations produced by CNNs, and add ancillary nets at the image level and the instance level respectively to prevent data distribution distortion caused by adversarial training. The experiments show that our proposed approach not only enhances the robustness of object detection under different environmental conditions, but also improves the detection accuracy of objects.