Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Inconsistency-based Active Learning for LiDAR Object Detection

Esteban Rivera, Loic Stratil, Markus Lienkamp

TL;DR

This paper tackles the labeling bottleneck in LiDAR-based 3D object detection for autonomous driving by introducing inconsistency-based active learning. It develops a LiDAR-specific sampling strategy using horizontal mirroring and two inconsistency scores, notably the Number-of-boxes inconsistency (NoB), to identify informative unlabeled samples. Experiments on KITTI with the PointPillars detector show that NoB-based sampling, particularly with an ascending ordering and retraining, can achieve the same mAP as random sampling with only 50% of the labeled data, with additional gains for underrepresented classes like pedestrians and cyclists. The findings suggest that simple, LiDAR-focused inconsistency metrics can outperform IoU-based measures, offering practical data-labeling efficiency and potential integration into future 3D detection training pipelines.

Abstract

Deep learning models for object detection in autonomous driving have recently achieved impressive performance gains and are already being deployed in vehicles worldwide. However, current models require increasingly large datasets for training. Acquiring and labeling such data is costly, necessitating the development of new strategies to optimize this process. Active learning is a promising approach that has been extensively researched in the image domain. In our work, we extend this concept to the LiDAR domain by developing several inconsistency-based sample selection strategies and evaluate their effectiveness in various settings. Our results show that using a naive inconsistency approach based on the number of detected boxes, we achieve the same mAP as the random sampling strategy with 50% of the labeled data.

Inconsistency-based Active Learning for LiDAR Object Detection

TL;DR

This paper tackles the labeling bottleneck in LiDAR-based 3D object detection for autonomous driving by introducing inconsistency-based active learning. It develops a LiDAR-specific sampling strategy using horizontal mirroring and two inconsistency scores, notably the Number-of-boxes inconsistency (NoB), to identify informative unlabeled samples. Experiments on KITTI with the PointPillars detector show that NoB-based sampling, particularly with an ascending ordering and retraining, can achieve the same mAP as random sampling with only 50% of the labeled data, with additional gains for underrepresented classes like pedestrians and cyclists. The findings suggest that simple, LiDAR-focused inconsistency metrics can outperform IoU-based measures, offering practical data-labeling efficiency and potential integration into future 3D detection training pipelines.

Abstract

Deep learning models for object detection in autonomous driving have recently achieved impressive performance gains and are already being deployed in vehicles worldwide. However, current models require increasingly large datasets for training. Acquiring and labeling such data is costly, necessitating the development of new strategies to optimize this process. Active learning is a promising approach that has been extensively researched in the image domain. In our work, we extend this concept to the LiDAR domain by developing several inconsistency-based sample selection strategies and evaluate their effectiveness in various settings. Our results show that using a naive inconsistency approach based on the number of detected boxes, we achieve the same mAP as the random sampling strategy with 50% of the labeled data.
Paper Structure (18 sections, 2 equations, 8 figures, 1 table)

This paper contains 18 sections, 2 equations, 8 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related work
  3. Active Learning
  4. Inconsistency analysis
  5. 3D active learning
  6. Method
  7. Active Learning cycle
  8. Inconsistency definition
  9. Experiment
  10. Model
  11. Results
  12. Inconsistency ordering
  13. Active learning cycle
  14. Class proportion
  15. Inconsistency proportion
  16. ...and 3 more sections

Figures (8)

  • Figure 1: Point cloud and its augmented version side to side, with white colored boxes representing objects which where consistently recognised with a 3d Detector in both point clouds, and a red box representing one car which was recognised only in one point cloud. This is an example of a point cloud which is interesting to label, because the model lacks robustness in its representation. Image from the KITTI dataset Geiger2012CVPR.
  • Figure 2: Pseudo active learning cycle results. Deviation for the baseline is shown as the blue area
  • Figure 3: Results for Pedestrian
  • Figure 4: Results for Cyclist
  • Figure 5: Results for Car
  • ...and 3 more figures