Effect of Encoder Architectures on the Generation of Vector Representations for Modeling 3D Objects via the Space of Convex Sets

This paper presents a study on the effect of different encoder architectures on the generation of vector representations for 3D object models. The central model is an autoencoder: the encoder generates a vector representation, and the vector-based decoder reconstructs the 3D object through the space of convex sets. While many studies focus on decoder optimization, this work investigates the impact of different encoder architectures on the quality of 3D model generation, using a novel data format. Additionally, we propose a new method for training encoders, involving a modified error function to improve the representation of generated vectors. Our experiments demonstrate that the ResNet50 architecture combined with the proposed loss function outperforms other architectures and methods in terms of Chamfer Distance (CD) and Light Field Distance (LFD) metrics. Specifically, the ResNet50 with the new loss function achieved a CD of 1.41 and an LFD of 3794.59, which are better and lower than the CD of 1.58 and LFD of 3866.41 produced by the DT-Net system, showing improvements of 0.17 in CD and 71.82 in LFD. These results suggest that deeper networks with sophisticated loss functions improve 3D model generation.