Teachers in concordance for pseudo-labeling of 3D sequential data

TL;DR

This paper introduces Concordance, a multi-teacher pseudo-labeling method leveraging temporal information in 3D point cloud sequences to improve semi-supervised learning in autonomous driving scenarios.

Contribution

It proposes a novel teacher ensemble with confidence-guided pseudo-labeling for 3D data, outperforming standard methods with minimal manual labels.

Findings

Outperforms some fully supervised methods with only 20% manual labels.

Significant improvement for rare classes in training data.

Effective in 3D semantic segmentation and object detection.

Abstract

Automatic pseudo-labeling is a powerful tool to tap into large amounts of sequential unlabeled data. It is specially appealing in safety-critical applications of autonomous driving, where performance requirements are extreme, datasets are large, and manual labeling is very challenging. We propose to leverage sequences of point clouds to boost the pseudolabeling technique in a teacher-student setup via training multiple teachers, each with access to different temporal information. This set of teachers, dubbed Concordance, provides higher quality pseudo-labels for student training than standard methods. The output of multiple teachers is combined via a novel pseudo label confidence-guided criterion. Our experimental evaluation focuses on the 3D point cloud domain and urban driving scenarios. We show the performance of our method applied to 3D semantic segmentation and 3D object detection on three benchmark datasets. Our approach, which uses only 20% manual labels, outperforms some fully supervised methods. A notable performance boost is achieved for classes rarely appearing in training data.