Semi-MapTR: Semi-Supervised Learning for End-to-End Online HD Map Construction

Abstract

Recently, the emergence of online high-definition (HD) mapping solutions for autonomous driving has shown great promise, such as MapTR and VectorMapNet. However, despite their success, these models depend heavily on large-scale annotated data for their training, thereby limiting their scalability due to the extensive manual labor required for data annotation. This thesis addresses this critical challenge by investigating the integration of semisupervised learning (SSL) into the online HD map construction framework. By leveraging a small portion of annotated data together with a larger portion of unlabeled data, the study aims to enhance the scalability and and efficiency of HD map generation. Therefore, we propose a semi-supervised approach built upon MapTRv2, Semi-MapTR, utilizing pseudo ground truth generated by rasterizing vectorized output of unlabeled data and enhancing confidence through occupancy grid mapping. We demonstrate an increase in both mean Average Precision (mAP) and mean Intersection over Union (mIoU) for both camera and LiDAR (Light Detection and Ranging) data using Semi-MapTR compared to MapTRv2, when both models are trained on the same amount of labeled data.