Year: 2025
Type: article
Abstract: 3D semantic segmentation, the process of assigning semantic labels to every point in a 3D space, is critical for numerous applications, including autonomous driving, robotics, medical imaging, and urban mapping. Despite significant progress, challenges such as data imbalance, scalability, and real-time processing constraints persist. This study addresses the real-time processing issue by comparing Tiny, Medium, and Large PointNet-inspired models utilizing the ShapeNetCore dataset. The models incorporate the T-Net module for pose normalization to maintain robustness against geometric transformations. Class-specific segmentation is explored by training separate models for the Airplane, Motorbike, and Car classes, allowing custom optimizations for each class. The Tiny model with 512 sampled points where the batch size is 16 and trained for 40 epochs with a starting learning rate of 1×10^(-3) achieved an average training accuracy of 86.18% and an average validation accuracy of 83.50%, making it optimal for real-time applications due to its fast inference speed and high accuracy.
Authors: Furkan Karaman, Fatma Gümüş
Institution: Milli Savunma Üniversitesi
Language: English
Subfield: Computational Mechanics
Field: Engineering
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W4408784180