A generative target detection algorithm based on diffusion model

Inspired by the Diffusion Model, this paper proposes a Self-attention-generating object detection algorithm (SAG-D). The algorithm uses self-attention to extract features and fuse them into the backbone network, so as to obtain more effective features to guide the decoder to find the appropriate RoI, and uses the principle of diffusion model to continuously correct the simulation box with noise to generate the predicted ground truth box. In the end, The experiments show that the proposed algorithm achieves very good results on MS COCO dataset.