Assessing the Impact of Image Quality on Multi-Task Learning Models for 3D Object Detection and Drivable Area Segmentation
Year: 2025
Type: conference-paper
Abstract: In autonomous driving, the quality of input images is crucial for the accuracy and reliability of perception systems, particularly in tasks like 3D object detection and drivable area segmentation.This study examines the impact of training Multi-Task Learning (M-TL) models exclusively on high-quality images for these applications.Using the KITTI dataset, we apply AI-based and traditional Image Quality Assessment (IQA) algorithms to filter and retain only high-quality images during training.Our experiments reveal that models trained on highquality images achieve significantly better performance than those trained on the full dataset, including images of varying quality.These findings highlight the critical role of image quality in enhancing the accuracy and robustness of M-TL learning models for autonomous driving.Furthermore, this work emphasizes the importance of integrating image quality evaluation into the data-preprocessing pipeline to optimize model performance.
Source: Computer Science Research Notes
Authors: Firas Jendoubi, Redouane Khemmar, Romain Rossi, Madjid Haddad, Segula Technologies
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W4412762974