DSP R-CNN: Direct Set Prediction Region with CNN features

Convolutional Neural Network (CNN) based object detection models struggle with differentiating objects from the background and with the separation and global interaction among multiple objects within an image. As a result, accurately capturing the location of objects in such images necessitates the use of the Region Proposal Network (RPN) structure. However, RPNs present several challenges in terms of performance and efficiency. This situation has led to an increasing focus on research in Transformer-based object detection models. While these Transformer-based models improve performance over their predecessors, they often compromise efficiency in terms of speed and training duration. The proposed method introduces a novel approach that interprets the channels of the feature map as compressed objects, fundamentally transforming the CNN paradigm by eliminating the need for Region Proposal in CNN-based object detection architectures. Utilizing a one-to-one matching function, it turns object detection into a direct prediction problem. Moreover, the DSP R-CNN model, developed from this method, streamlines the pipeline by dispensing with heuristic elements like Non-Maximum Suppression (NMS) and anchor box generation. The experiments on Circular pipe dataset show that this approach achieves higher accuracy and faster performance compared to the widely used CNN-based model Faster R-CNN and Transformer-based model DETR in the field of object detection.