Improved Deep Learning Architecture For Depth Estimation

摘要

Numerous benefits of depth estimation from the single image field on medicine, robot video games and 3D reality applications have garnered attention in recent years. Closely related to the third dimension of depth, this operation can be accomplished using human vision, though considered challenging due to the various issues when using computer vision. The differences in the geometry, the texture of the scene, the occlusion scene boundaries and the inherent ambiguity exist because of the minimal information that could be gathered from a single image. This paper, therefore, proposes a novel depth estimation in the field of architecture, which includes the stages that can manage depth estimation from a single RGB image. An encoder-decoder architecture has been proposed, based on the improvement yielded from DenseNet that extracted the map of an image using skip connection technique. This paper also takes on the reverse Huber loss function that essentially suits our architecture hand driven by the value distributions that are commonly present in depth maps. Experimental results have indicated that the depth estimation architecture that employs the NYU Depth v2 dataset has a better performance than the other state-of-the-art methods that tend to have fewer parameters and require fewer training time.