Abstract
We propose DiffusionDet, a new framework that formulates object detection asa denoising diffusion process from noisy boxes to object boxes. During thetraining stage, object boxes diffuse from ground-truth boxes to randomdistribution, and the model learns to reverse this noising process. Ininference, the model refines a set of randomly generated boxes to the outputresults in a progressive way. Our work possesses an appealing property offlexibility, which enables the dynamic number of boxes and iterativeevaluation. The extensive experiments on the standard benchmarks show thatDiffusionDet achieves favorable performance compared to previouswell-established detectors. For example, DiffusionDet achieves 5.3 AP and 4.8AP gains when evaluated with more boxes and iteration steps, under a zero-shottransfer setting from COCO to CrowdHuman. Our code is available athttps://github.com/ShoufaChen/DiffusionDet.