Unsupervised Monocular Depth Estimation Based On Convolutional Neural Network

Accurate depth estimation is a foundation that computer can understand and sense the real 3-D world and has many important applications in the field of augmented reality,3-D reconstruction,autonomous driving,etc.However,it’s very expensive that acquiring depth by means of hardware facility.Therefore,more and more researchers launch into a study that inferring depth from a 2-D image by use of deep learning.Supervised monocular depth estimation needs expensive depth labels during training,while Unsupervised monocular depth estimation can be trained with image sequences and is greatly worth studying.However,there still exists many problems about unsupervised monocular depth estimation and large gaps with supervised monocular depth estimation.For existing problems in unsupervised monocular depth estimation,we launch into such researches:(1)For depth information loss of objects with few pixels in a scene and ambiguous boundary because of semantic gaps,we improve network structure.Proposed module of full-scale feature aggregation can successively and progressively aggregate features along layers that is efficient to decrease loss of feature pixels during the transmission process of feature maps and preserve more full depth information of prediction results.In the decoder phase,the aggregated features are fused by efficient channel attention mechanism,that can decrease gaps between low-resolution semantic information and high-resolution spatial features and guides generation of sharp boundary by unsupervised monocular depth estimation.Experiments on the KITTI datasets show that the proposed methods preserve more depth information,produce shaper boundary and improve the prediction accuracy of depth estimation.(2)For large mistakes in the depth discontinuous regions because the local receptive field of basic convolution can’t model relationships between long-range pixels from a scene,we propose adaptive global feature enhance module by means of self-attention mechanism,that improves smoothness of depth predictions by global feature enhance for high-level semantic feature.And stimulating performance potential of proposed model by adding depth hints in order to relieve low quality problems for depth maps because of instability of loss function.Experiments on the KITTI datasets show that the proposed methods can produce smoother depth results and decrease the error of depth prediction.Figure [38] Table [11] Reference [64]...