Learning Depth for Multi-View Stereo with Adversarial Training
Liang Wang, Deqiao Fan, Jianshu Li · 2021
The deep learning-based Multi-View Stereo (MVS) methods have shown excellent performance in depth estimation. However, due to the heavy computational burden caused by 3D convolution, most of the existing methods are difficult to be applied to high-resolution scenarios. In this paper, a novel method of learning depth for multi-view stereo with adversarial training is proposed, which exploits the Wasserstein generative adversarial network (GAN) to overcome shortcomings of the original GAN. The proposed network mainly consists of the encoder, the generator, and the discriminator sub-network. With the help of the proposed loss function, the proposed network can merge more global information provided by the discriminator of GAN and local information provided by the generator. This can capture more details and estimate an accurate depth map in high-resolution scenarios. It also significantly reduces memory consumption by exploiting WGAN instead of performing 3D convolution. Extensive experiments validate the proposed method, which can achieve on par or even better performance than the state-of-the-art.