This academic paper introduces PillarNeSt, a new technique for 3D object detection in point clouds. The authors highlight the effectiveness of 2D backbone scaling and pretraining for pillar-based 3D object detectors, which have traditionally relied on randomly initialized 2D convolution neural network (ConvNet) for feature extraction. Using ConvNets that are pretrained on large-scale image datasets, the authors demonstrate that PillarNeSt outperforms existing 3D object detectors on nuScenes and Argoversev2 datasets. The paper discusses the challenges and possible solutions for scaling up and transferring image knowledge to point cloud data.
Publication date: 30 Nov 2023
Project Page: Not Provided
Paper: https://arxiv.org/pdf/2311.17770
