VoloGAN: Adversarial Domain Adaptation for Synthetic Depth Data
This addresses the data scarcity problem for 3D reconstruction algorithms by enabling synthetic-to-real domain adaptation with few training samples, which is incremental as it builds on existing CycleGAN frameworks.
The paper tackles the problem of generating realistic training data for single-view 3D reconstruction by translating synthetic RGB-D images from high-quality 3D models into images that mimic consumer depth sensors, using a CycleGAN-based adversarial domain adaptation network called VoloGAN.
We present VoloGAN, an adversarial domain adaptation network that translates synthetic RGB-D images of a high-quality 3D model of a person, into RGB-D images that could be generated with a consumer depth sensor. This system is especially useful to generate high amount training data for single-view 3D reconstruction algorithms replicating the real-world capture conditions, being able to imitate the style of different sensor types, for the same high-end 3D model database. The network uses a CycleGAN framework with a U-Net architecture for the generator and a discriminator inspired by SIV-GAN. We use different optimizers and learning rate schedules to train the generator and the discriminator. We further construct a loss function that considers image channels individually and, among other metrics, evaluates the structural similarity. We demonstrate that CycleGANs can be used to apply adversarial domain adaptation of synthetic 3D data to train a volumetric video generator model having only few training samples.