Mirror R-CNN: Object Detection with Flipped Image

Through the study of the sparse query-based detector, we find that detecting the flipped images instead of original images could improve the detection accuracy on many widely used dataset. Based on this phenomenon, a novel Sparse R-CNN based detector named as Mirror R-CNN is proposed in this paper, which contains the mirror processing module and confidence selection module. The mirror processing module is used to feed the flipped images into the detection network and mapped the detection results of the flipped images back to the original images. While the final detection results are obtained based on the confidence scores of the original and flipped images through confidence selection module to improve the detection failure due to object deformation. The whole processing can be viewed as the data enhancement at the output side. Besides, the dynamic weight loss is proposed to relax the errors caused by the manual data annotation and increase the convergence rate in training process. The proposed Mirror R-CNN could obtain comparable results without significantly increasing the computation. On the normally used datasets, such as CrowdHuman and COCO, Mirror R-CNN could get an average 0.3% improvement in AP and 1% improvement in APs.