Monocular Depth Estimation Method Based on Dual-discriminator Generative Adversarial Networks

To solve the problem of the low accuracy of self-supervised monocular depth estimation, a self-supervised monocular depth estimation method based on dual-discriminator generative adversarial networks were proposed in this paper. The advantages of generative adversarial networks were adopted to synthesize visually credible images, and further improve the accuracy of self-supervised monocular depth estimation. First, to make full use of the reconstructed image, the Wasserstein generative adversarial networks were improved and the structure of two discriminators was constructed. The dual-discriminator had more stringent requirements for the generator and training objectives, avoiding the information loss caused by the introduction of discriminator only on the left image or right image. Second, according to the structure of the network, a local-global consistent loss function was proposed to ensure the authenticity of pixels and the consistency of local-global content. Results show that the proposed method effectively improves the accuracy of monocular depth estimation and has better performance of depth estimation.