Abstract
Object detection, a quintessential task in the realm of perceptual computing,can be tackled using a generative methodology. In the present study, weintroduce a novel framework designed to articulate object detection as adenoising diffusion process, which operates on the perturbed bounding boxes ofannotated entities. This framework, termed \textbf{ConsistencyDet}, leveragesan innovative denoising concept known as the Consistency Model. The hallmark ofthis model is its self-consistency feature, which empowers the model to mapdistorted information from any time step back to its pristine state, therebyrealizing a \textbf{``few-step denoising''} mechanism. Such an attributemarkedly elevates the operational efficiency of the model, setting it apartfrom the conventional Diffusion Model. Throughout the training phase,ConsistencyDet initiates the diffusion sequence with noise-infused boxesderived from the ground-truth annotations and conditions the model to performthe denoising task. Subsequently, in the inference stage, the model employs adenoising sampling strategy that commences with bounding boxes randomly sampledfrom a normal distribution. Through iterative refinement, the model transformsan assortment of arbitrarily generated boxes into definitive detections.Comprehensive evaluations employing standard benchmarks, such as MS-COCO andLVIS, corroborate that ConsistencyDet surpasses other leading-edge detectors inperformance metrics. Our code is available athttps://anonymous.4open.science/r/ConsistencyDet-37D5.