Fused voxel autoencoder for single image to 3D object reconstruction

Abstract

The encoder-decoder based models become popular for many computer vision problems from image generation to image segmentation tasks. Due to the impressive performance of these models and Generative Adversarial Networks on image generation, these models have also adopted to 3D domains for 3D reconstruction and generation tasks. In this Letter, the authors have also attempted to solve a single image to 3D reconstruction problem by a novel encoder-decoder based model that is based on a fusion of encoders with the weaksupervision approach. Thus, they have focused on multi-category models inspiring the outstanding single-category models on literature for real-world tasks. In the experiments, it is seen that the proposed model is capable of generating the 3D objects from a single image by benefiting a wide range of features with weak class annotations.