Depth Estimation of Monocular Images using Transfer Learning based Unet Model

Abstract

The problem of depth estimation using monocular images is very challenging contrasted to methods for estimating depth that involve several pictures, like stereo depth perception. Previous studies in this field have usually focused on utilizing geometrical priors or relied on other data collection techniques. Various machine learning methods, notably deep convolutional neural networks (CNN) integrated with artificial intelligence (AI) approaches, have recently produced new marks for a variety of visual applications. In this paper, a convolution neural network is used for estimating high-resolution depth image. We have used a pretrained model DenseNet-169 which is trained on ImageNet. The encoder-decoder model used is a simple Unet model, used in biomedical image analysis. The proposed model is more accurate and efficient with reduced complexity in terms of the fewer parameters used for training. This model is also noteworthy when comparing a state of art and qualitatively it performs well and captures better edges and corners of the depth map, which is the most important factor in the depth estimation.