Diffusion-based dynamic super-dense candidate boxes with random center points for 3D object detection

Abstract

Diffusion models have achieved promising results in image generation, but their applications in 3D object detection still need further exploration. In this paper, we design a novel model DiffCandiDet based on dense heads with Gaussian distributed center points for 3D object detection, which effectively integrates the anchor-based method and the Gaussian random noise-based method to leverage the powerful denoising and reconstruction capabilities of the diffusion model. To achieve the learning balance for multi-class 3D object detection, we propose a Dynamic Super-dense Candidate Boxes (DSCB) strategy. Notably, DiffCandiDet addresses the issue of traditional models struggling to detect pedestrians walking side by side. In addition to Gaussian distribution, we also propose a DSCB strategy based on discrete uniform distribution (DUCandiDet) and continuous uniform distribution (CUCandiDet), to reduce the runtime consumption and enhance the robustness of the model. Extensive experiments show that DiffCandiDet achieves competitive results on both KITTI and Waymo Open Datasets. DiffCandiDet ranks 1st on the KITTI validation set in the Car and Pedestrian detection leaderboard. Code is available at https://github.com/SiHengHeHSH/DiffCandiDet.