Adaptive threshold cascade faster RCNN for domain adaptive object detection

Abstract

Object detection usually assumes that training and test data come from the same distribution, but the assumption is not always hold in practice. Due to domain shift problem, applying a trained detector to a new domain will lead to a great decrease in detection accuracy. Domain adaptive object detection has been adopted to maintain high detection accuracy in the face of various domain shift problems. Domain adaptive object detection methods mainly include adversarial-based methods, discrepancy-based methods, reconstruction-based methods, hybrid methods and others. Domain adaptive Faster RCNN is a classical adversarial-based method. In order to further improve the accuracy of domain adaptive object detection, we propose a method based on the Domain adaptive Faster RCNN called adaptive threshold cascade Faster RCNN (ATCFR). The ATCFR introduces the cascade strategy and adaptive threshold strategy. The cascade strategy improves the quality of bounding boxes and solves the problem of overfitting and mismatch in Faster RCNN. The adaptive threshold strategy ensures the balance of positive and negative samples and we don’t have to manually set the threshold as we did in cascade RCNN. In the end, we evaluate our new approach by using four classic datasets, including Cityscapes, Foggy Cityscapes, SIM 10k and KITTI. Experimental results show that our method has higher accuracy in variousdomain shift problems, compared with the state-of-the-art methods.