DPM-Det: Diffusion Model Object Detection Based on DPM-Solver++ Guided Sampling

DiffusionDet is the first neural network model to apply the diffusion model to object detection, which is a new object detection paradigm that achieves good results compared to other mature object detection models. However, the diffusion model is an iterative model by nature, which requires setting numerous sampling steps to get good results. DiffusionDet has difficulty obtaining stable samples with fewer sampling steps, and calling the detection head once per iteration will cause the inference time to keep increasing. To alleviate this problem, the DPM-Det algorithm is proposed, which only changes the sampling process from noise boxes to bounding boxes in progressive refinement during inference, while keeping the structure of the DiffusionDet model unchanged. It replaces the original denoising diffusion implicit models in DiffusionDet by introducing DPM-Solver++. On the MS-COCO dataset, with ResNet-50 with FPN as the backbone and the training and evaluation boxes fixed at 500, DPM-Det achieved an AP of 47.0 with 5 steps and an AP of 47.3 with 8 steps. Compared with DiffusionDet, DPM-Det has a certain degree of improvement in both detection accuracy and speed, which can achieve a better balance of accuracy and speed.