PointPillars: Fast Encoders for Object Detection from Point Clouds, Lang, Vora, Caesar, Zhou, Yang, Beijbom; 2018 - Summary
author: aabayomi
score: 8 / 10

Core Idea

The authors proposes faster and efficient approach for 3D object detection and representation. PointPillars is an encoder that learns 3D point cloud representations by stacking learned features as vertical tensors and flattens to 2D pseudo image.

Main Contributions

PointPillars can be used with an standard 2D backbone.

Loss function

The loss formulation combines the sotfmax classification loss, focal loss on object classification and localization loss.This is optimized using Adam of learning rate of $2 * 10^−4$ with a decay rate 0.8 at every 15 epoch.

Data augmentation was crucial to achieving great performance and one of the approaches used was the random selection of ground truth samples from a lookup table.


On KITTI datasets PointPillars outperforms fusion based methods recognizing cars and cyclists with respect to mean average precision (mAP) compared to lidar-only methods.

PointPillar showed better accuracy and speed compared to SOTA methods while varying the size of the spatial binning.