Unsupervised Learning of 3D Object Reconstruction with Small Dataset

We propose an unsupervised learning framework trained with a small dataset for 3D object reconstruction from a single image. Our method utilizes autoencoders to extract 3D knowledge from an image, a differentiable renderer to generate an image from a reconstructed 3D object, and GAN inversion to produce pseudo images with random viewpoints and lighting to enlarge the training dataset. Quantitative and qualitative experimental results prove that our approach can recover 3D shapes with small dataset as accurately as state-of-the-art networks with large dataset. ? 2021 IEEE