CRNet: Combining CenterNet and R-CNN for Object Detection in Traffic Scenes
Zeyu Cui, Jun Yu · 2023
Object detection is typically formulated by researchers as a bounding box regression task(the two-stage method) or an object center localization task (the one-stage method). However, the methods incorporating the above two ideas make inferior performance in terms of both inference speed and detection accuracy, with quite limited applications in real-world traffic scenes. In this paper, we propose a simple but effective method named Center-RCNN(CRNet) which incorporates the aforementioned two ideas. Specifically, we first adopt the methods of object center localization instead of typical region proposal network(RPN) to conduct proposal selection in a fast and simple way, and then use the bounding box regression method to predict the final object location. The whole network combines the advantages of the two methods to keep a balance between inference accuracy and speed. Experiments are conducted on the BDD100K dataset, and the results prove the effectiveness of our method.