A depth estimation method for multi view and high precision images

• 1、Intelligence Science and Technology Center of BUPT，Beijing 100876

Abstract：Aiming at the problem that high-precision images are difficult to reconstruct effectively in dense multi view reconstruction, a learning-based depth estimation method is proposed. In this method, the dilated convolution neural network iss used to extract image features, and the Long Short-Term Memory network is applied to construct and optimize the cost volume. Besides, the supervised and unsupervised training methods are adopted. Experimental results on two real scene multi view image datasets show that the proposed method not only outperforms state-of-the-arts, but also is several times less in GPU memory application in comparison with traditional methods and other learning-based methods. Therefore, it can be used for reconstruction of high-precision images while the accuracy and integrity of model depth prediction are improved.

Keywords： deep learning convolution neural network multi view reconstruction