ADAPTIVE WEIGHTED IOU-FOCAL (AWIF) LOSS FUNCTION FOR ENHANCED 3D OBJECT GENERATION: A COMPARATIVE STUDY USING 3D-VAE-GAN ARCHITECTURE

This study evaluates the efficacy of a three-dimensional (3D) Variational Autoencoder-Generative Adversarial Network (3D-VAE-GAN) architecture in generating objects from images, with a focus on reconstruction loss functions. The analysis employs several loss functions, including Mean Squared Error (MSE), Binary Cross-Entropy (BCE), Focal, Dice, Intersection Over Union (IoU), and the newly developed Adaptive Weighted IoU-Focal (AWIF), to compare the effect of the metrics. In experiments conducted with 13 different objects from the ShapeNet dataset, both quantitative and qualitative assessments indicated that the AWIF loss function outperformed other loss functions. While volumetric overlap metrics (IoU) showed saturation due to the inherent sparsity of voxel grids, the proposed AWIF loss significantly improved surface geometric fidelity. Specifically, it reduced the Chamfer Distance (CD) error by 7.3% compared to MSE and 10.2% compared to standard IoU loss, proving its superior capability in reconstructing fine surface details. Furthermore, time analysis of the AWIF loss function was conducted, demonstrating its potential competitiveness in this domain.