Paper Image

Pillar-based 3D detection with scaled and pretrained backbones

Published on:

29 November 2023

Primary Category:

Computer Vision and Pattern Recognition

Paper Authors:

Weixin Mao,

Tiancai Wang,

Diankun Zhang,

Junjie Yan,

Osamu Yoshie

Bullets

Key Details

PillarNeSt backbones are specially designed for point cloud sparsity/irregularity, with rules like large kernels

Scaling model size consistently improves detection performance, unlike prior work

ImageNet pretraining accelerates convergence and boosts accuracy

PillarNeSt sets new state-of-the-art on nuScenes and Argoverse datasets

AI generated summary

Pillar-based 3D detection with scaled and pretrained backbones

This paper explores using scaled-up and ImageNet-pretrained convolutional backbones to improve pillar-based LiDAR 3D detection, embracing techniques from the image domain. The authors design backbones tailored to point cloud properties, then scale model capacity and initialize weights from pretrained ConvNeXT models. Their method, PillarNeSt, substantially outperforms prior work on nuScenes and Argoverse datasets.

Answers from this paper

Comments

No comments yet, be the first to start the conversation...

Sign up to comment on this paper

Sign Up