Dual-Attention Sparse R-CNN via Single ROI Transformer and Dynamic CBAM

A critical drawback of CNN-based object detection is an additional post processing such as non-maximum suppression(NMS). Recent works have solved that problem by applying a vision transformer, which is able to make sparse correlation between feature maps and proposal embedding. Sparse R-CNN, which is end-to-end object detection network, applies dynamic convolution for the correlation but requires many parameters. Based on the sparse R-CNN, this paper presents a dual-attention module for accurate and efficient end-to-end object detection. The proposed dual-attention method consists of: i) dynamic CBAM, which is more efficient than dynamic convolution and ii) single ROI transformer. Experimental results show that the proposed method not only is efficient in the sense of reduced number of parameters but also improves the performance.