Pillar-based Object Detection for Autonomous Driving

TL;DR

This paper introduces a pillar-based, anchor-free 3D object detection framework optimized for autonomous driving, leveraging cylindrical projection and pillar-to-point alignment to improve accuracy and simplicity over previous methods.

Contribution

It presents a novel pillar-based, anchor-free detection approach that simplifies 3D object detection and enhances performance in autonomous driving scenarios.

Findings

Significantly outperforms state-of-the-art methods

Reduces hyperparameter tuning complexity

Improves detection accuracy in sparse point clouds

Abstract

We present a simple and flexible object detection framework optimized for autonomous driving. Building on the observation that point clouds in this application are extremely sparse, we propose a practical pillar-based approach to fix the imbalance issue caused by anchors. In particular, our algorithm incorporates a cylindrical projection into multi-view feature learning, predicts bounding box parameters per pillar rather than per point or per anchor, and includes an aligned pillar-to-point projection module to improve the final prediction. Our anchor-free approach avoids hyperparameter search associated with past methods, simplifying 3D object detection while significantly improving upon state-of-the-art.