Find in Library

Depth computation from an image is useful for many robotic systems like obstacle recognition, autonomous navigation, and 3D measurements. The estimation is best solved with Deep Neural Networks (DNN) as these are non-linear and ill-posed problems. The network takes single-color images with corresponding ground truth to predict depth-map after training. The depth accuracy, here, is dependent on the quality of ground truth and training images. Images have inherent blurs, which impact depth prediction and accuracy. In our work, we study different combinations of loss functions involving various edge functions to improve the depth of images. We use DenseNet and transfer learning method for learning and prediction of depth. Our analysis shows improvement in performance parameters as well as in the visual depth-map. We achieve 85% δ1 accuracy and improve log10error using NYU Depth V2 dataset.