Research on the Optimization of the Detr Model Based on Hybrid Supervision in Object Detection

Object detection is a key technology in the field of computer vision. It can identify and locate specific objects in images and is widely used in many fields. DETR is an end-to-end object detection method based on the Transformer architecture. Compared with traditional methods, it can simplify the application and deployment process and has better performance and scalability. However, its training and convergence are slow and the cost is high. This paper proposes the DETR hybrid supervision algorithm. Firstly, based on the hybrid supervision training method, negative samples are mined to improve the training speed and detection accuracy of DETR. Secondly, the ranking-aware dual-branch method utilizes the features of different attention layers for supervision and introduces losses to enhance the effect. Thirdly, the above methods are combined with PETR and applied to 3D detection. Experiments on mainstream 3D object detection datasets have verified the effectiveness of this method in 3D detection, solved the problem of slow training of DETR, and improved its performance.