Real-Time Rotated Object Detection Using Angle Decoupling

Real-time object detection and angle fitting of rotating objects in industrial scenes are challenging computer vision tasks at present. Current object-detection models in natural scenes cannot meet the needs of the industry for accurate angle fitting. Thus, we propose a rotating detection box representation based on angle information decoupling (Rbbox), and designed rotated-YOLO (You Only Look Once) deep learning rotating object detection model for Rbbox representation. First, an angle information encoding and decoding method is proposed by analyzing the problem of calculating angle error mutation in the existing angle information expression methods. Second, based on the YOLO series neural network model, the detection head and loss function are improved. Finally, the cosine similarity evaluation index was proposed to solve the problem in which intersection over union could not clearly express the accuracy of angle fitting in the object detection task in natural scenes. Experiments on industrial datasets show that the algorithm performs well on open-source industrial datasets. The average angle cosine similarity between rbbox and label box is 0.98781, and mAP@.5 is 0.995. The reasoning speed of NVIDIA A100 GPU is 76.923 fps, which can meet the needs of real-time reasoning in industry.