Original scientific paper

https://doi.org/10.17559/TV-20251125003159

Entropy-Maximized Generative Adversarial Networks for Point Cloud Data Generation and Augmentation in Machine Vision

Chuang Li ; School of Electronic and Information Engineering, Liuzhou Polytechnic University, Liuzhou 545006, China
Jian Yu ; School of Electronic and Information Engineering, Liuzhou Polytechnic University, Liuzhou 545006, China *

* Corresponding author.

Abstract

This work presents EM-GAN (Entropy-Maximized Generative Adversarial Network), designed for point cloud generation to confront the enduring difficulties of 3D structure reconstruction. In comparison with established baselines such as PC-GAN and PUFA-GAN, the method delivers notable gains in sample diversity, geometric fidelity, and reconstruction precision, while more effectively exploiting the latent feature representation. To alleviate mode collapse, the framework integrates an entropy-guided regularization component that stimulates broader output variability. Its architecture combines a multi-stage generator, which incrementally enriches fine structural details, with a PointNet-based discriminator that promotes realistic 3D formations. As a result, EM-GAN produces point clouds that are more complete and accurate than those obtained from conventional counterparts. Experimental validation on the ShapeNet dataset shows consistent improvements across widely recognized indicators, including Inception Score (IS), Fréchet Point Cloud Distance (F-PointNet), and Chamfer Distance (CD), thereby substantiating both its reconstruction quality and latent-space utilization. Further qualitative analysis underscores the framework's ability to synthesize high-resolution and perceptually faithful shapes. Overall, EM-GAN offers a robust paradigm for point cloud synthesis, contributing substantial progress in variety, accuracy, and representational scope, with subsequent research directed toward optimization and deployment on large-scale datasets and real-world applications.

Keywords

3D shape and object analysis; adversarial learning frameworks; computer vision applications; deep neural models; point cloud processing

Hrčak ID:

350425

URI

https://hrcak.srce.hr/350425

Publication date:

31.8.2026.

Visits: 0 *