Monocular depth estimation with multi-view attention autoencoder

Geunho Jung, Sang Min Yoon

Abstract: Depth map estimation from a single RGB image is a fundamental computer vision and image processing task for various applications. Deep learning based depth map estimation has improved prediction accuracy compared with traditional approaches by learning huge numbers of RGB-D images, but challenging issues remain for distorted and blurry reconstruction in object boundaries because the features are not enforced during training. This paper presents a multi-view attention autoencoder embedded in a deep neural network to emphasize self-representative features, which provide robust depth maps by simultaneously accentuating useful features and reducing redundant features to improve depth map estimation performance. Qualitative and quantitative experiments were conducted to verify the proposed network effectiveness, which can be utilized for three-dimensional scene reconstruction and understanding.