ATF-3D: Semi-Supervised 3D Object Detection With Adaptive Thresholds Filtering Based on Confidence and Distance | IEEE Journals & Magazine | IEEE Xplore

ATF-3D: Semi-Supervised 3D Object Detection With Adaptive Thresholds Filtering Based on Confidence and Distance


Abstract:

Performance of current point cloud-based outdoor 3D object detection relies heavily on large-scale high-quality 3D annotations. However, such annotations are usually expe...Show More

Abstract:

Performance of current point cloud-based outdoor 3D object detection relies heavily on large-scale high-quality 3D annotations. However, such annotations are usually expensive to collect and outdoor scenes easily accumulate massive unlabeled data containing rich scenes. Semi-supervised learning is a effective alternative to utilize both labeled and unlabeled data, but remains unexplored in outdoor 3D object detection. Inspired by indoor semi-supervised 3D detection methods, SESS and 3DIoUMatch, we propose ATF-3D, a semi-supervised 3D object detection framework for outdoor scenes. Specifically, we design a simple yet effective adaptive thresholds search method based on distances and categories for obtaining high-quality pseudo labels. Concurrently, we propose an iterative training mechanism with pseudo-label training and self-ensembling learning to combine the advantages of both schemes. Furthermore, we adopt point cloud data augmentations in the self-ensembling learning stage to further improve the performance. Our ATF-3D ranks first among all single-model methods in the ONCE benchmark. Results on both ONCE and Waymo datasets demonstrate substatial improvements over the supervised baseline.
Published in: IEEE Robotics and Automation Letters ( Volume: 7, Issue: 4, October 2022)
Page(s): 10573 - 10580
Date of Publication: 30 June 2022

ISSN Information:

Funding Agency:


References

References is not available for this document.