One-Stage Lightweight Network of Object Detection for Rectangular Panoramic Images

Nowadays, object detection has developed rapidly and the application scenarios of panoramic image detection are increasing. Compared with ordinary images, panoramic images have a certain degree of distortion of the objects and the number of objects is greater. Therefore, the traditional target detection network designed for vanilla images will bring problems such as insufficient feature extraction and slow inference speed of object detection. In this paper, we proposed a one-stage detection network to solve the problems. First, we constructed the ELAN-P module with Partial Convolution (PConv) to reduce the computational complexity. Second, we introduced the bi-level attention mechanism Biformer into the network to improve the robustness of the detection network and better capture the distortion information. Finally, we made a panoramic dataset to train the detection model and evaluate the performance of the proposed model. The experiment results verified the effectiveness of our model compared to the popular networks.