A Robust Cross-scene Object Detection Method

In the field of computer vision,object detection is one of the research focuses.With the rise of deep neural networks,many convolutional neural networks such as ALex Net,Faster R-CNN and YOLO have emerged one after another,which has greatly improved the performance of object detection based on large amount of annotated data.However,due to the existence of distribution discrepancy between different scenarios(domains),it still remains challenging to deploy a pre-trained object detector in new unseen domains,which is common in practical applications.Recently,many research efforts have been devoted to cross-domain object detection,which aims to solve the problem of how to generalize the pre-trained detector to a new unlabeled target domain data.At present,the common adversarial-based cross-domain object detection methods can only try to make the marginal distribution of the inputs from two domains to be aligned,however,can not make the joint distribution of inputs and outputs to be aligned.Another common cross-domain object detection methods which using self-training to generate pseudo-labels,can make the joint alignment of distributions for inputs and outputs from two domains to be aligned,but the selection of pseudo-labels is a bottleneck.This paper proposes a robust cross-scene object detection method uncertainty-aware model adaptation for cross-domain object detection,considering that:(1)the estimation and exploitation of model uncertainty in a new domain is critical for robust domain adaptation;and(2)the joint alignment of distributions for inputs(feature alignment)and outputs(self-training)is needed.Firstly,in order to predict the uncertainty from the classification and regression branches of the model,this paper establishes a Bayesian CNN-based framework for uncertainty estimation in object detection.Secondly,in order to select the reliable predictions as the pseudo-labels for the target domain data,an uncertainty-aware pseudo-labels selection algorithm is proposed in this paper.At the same time,in order to enhance the robustness of the model in the process of feature alignment,an uncertainty-guided feature alignment method is proposed in this paper.Finally,this paper devise a new training scheme,that is,joint feature alignment and self-training of the object detection model with pseudo-labels.Finally,this paper makes an experimental comparison and performance analysis bases on experiments on four cross-domain object detection benchmarks,and the results show that the proposed method achieves state-of-the-art performance compared with the existing methods.