Title
Monocular 3D Shape Estimation for Autonomous Driving
Authors
[ 1 ] Wydział Automatyki, Robotyki i Elektrotechniki, Politechnika Poznańska | [ 2 ] Instytut Robotyki i Inteligencji Maszynowej, Wydział Automatyki, Robotyki i Elektrotechniki, Politechnika Poznańska | [ SzD ] doctoral school student | [ P ] employee
Scientific discipline (Law 2.0)
[2.2] Automation, electronics, electrical engineering and space technologies
Year of publication
2026
Chapter type
chapter in monograph / paper
Publication language
english
Keywords
Abstract
EN Monocular 3D shape estimation is crucial for autonomous driving, enabling accurate vehicle pose estimation from single images. This paper presents a deep learning-based approach utilizing a Variational Autoencoder (VAE) and Graph Convolutional Networks (GCNs) to estimate dense 3D meshes of vehicles. Unlike conventional keypoint-based methods, the proposed approach reconstructs complete car shapes, ensuring robust pose estimation even under occlusions and varying lighting conditions. A pipeline is introduced where a Vision Transformer extracts image features, followed by a Shape Head predicting a latent vector, which the VAE decoder converts into a full 3D mesh. The Apollo- Car3D dataset is used for training and evaluation, demonstrating that the meshbased method achieves improved accuracy of keypoint detection, while maintaining high accuracy of pose estimation. Results highlight the effectiveness of dense mesh prediction that can serve for enhancing vehicle detection, tracking, and collision avoidance in autonomous driving systems.
Date of online publication
02.01.2026
Pages (from - to)
153 - 164
Ministry points / chapter
20
System created by Poznań University of Technology
and Poznan Supercomputing and Networking Center
Log in through eKonto to add to SIS