Evaluation and Improvements of Deep-Learning-Based High Definition Map Models

Thumbnail Image

Files

master_Zabeo_Alberto_2024.pdf (7.36 MB)

URL

Journal Title

Journal ISSN

Volume Title

School of Science | Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.

Authors

Date

2024-09-30

Department

Major/Subject

Data Science

Mcode

Degree programme

Master's Programme in ICT Innovation

Language

en

Pages

53

Series

Abstract

High Definition (HD) maps are vital for autonomous driving, offering precise environmental representations to aid navigation and decision-making. Traditionally constructed via offline SLAM methods, these maps faced scalability and maintenance challenges. Recent advances have shifted towards online, deep learning-based methods, enabling real-time HD Map generation with improved adaptability and scalability. Early approaches treated HD Map construction as a segmentation problem in the Bird's Eye View (BEV) space, requiring extensive post-processing. Subsequent end-to-end solutions, such as VectorMapNet and MapTR, improved efficiency by directly predicting vectorized map elements, reducing the need for post-processing. Despite these advancements, current evaluation methods for HD Map models often use random dataset splits, which fail to account for spatial and locality dependencies inherent in the data, potentially leading to data leakage and overly optimistic performance assessments. This study investigates the effectiveness of existing evaluation criteria and proposes a novel dataset splitting strategy to better gauge model generalization. By analyzing state-of-the-art models under this new approach, the study aims to uncover limitations of current practices and suggest improvements in both evaluation methods and model architectures. Key questions include the efficacy of default data splits, the potential benefits of alternative splitting strategies, and opportunities for enhancing model performance through both architectural and model-agnostic approaches.

Description

Supervisor

Laaksonen, Jorma

Thesis advisor

Malek-Mohammadi, Reza
Dabbaghchian, Saeed

Keywords

data splitting, polyline clustering, temporal aggregation, HD map, deep learning, feature extraction

Other note

Citation

Permanent link to this item

https://urn.fi/URN:NBN:fi:aalto-202411217262

Collections

[dipl] Perustieteiden korkeakoulu / SCI