Contrastive learning for LIDAR point cloud segmentation

Loading...
Thumbnail Image

URL

Journal Title

Journal ISSN

Volume Title

Perustieteiden korkeakoulu | Master's thesis

Department

Major/Subject

Mcode

SCI3095

Language

en

Pages

82+1

Series

Abstract

Three dimensional high-definition point clouds containing semantic information are crucial in several applications, such as autonomous driving. In urban traffic scenarios, these point clouds are known to be affected by extreme class imbalance, the large planar surfaces being the dominant features, such as the drivable road area, while traffic signs, for instance, appear only rarely, making the training of deep learning-based segmentation methods especially difficult. To this end, we create our own multi-spectral dense LIDAR data set and conduct an ablation study, the results of which highlight the importance of tackling class imbalance, as well as the usability of LIDAR reflectance values in addition to point coordinates, as evidentiated by the final model achieving 0.3792 mIoU, a 67% relative gain over the baseline method. We also experiment with two different LIDAR sensors operating on different wavelengths (905 nm and 1550 nm), the results of the experiments showing that the sensor with a shorter wavelength performed better, suggesting that objects and materials commonly faced in urban traffic scenarios might be easier to differentiate at this end of the spectrum. To address the burden of manual labeling, and the need for large-scale data sets to train such networks, we successfully adapt label-efficient supervised contrastive methods as well from the image processing literature to learn point cloud latent representations usable for the downstream task of point cloud semantic segmentation.

Description

Supervisor

Ilin, Alexander

Thesis advisor

Kukko, Antero
Eichhardt, Iván

Other note

Citation