Novel Algorithm for Depth Map Estimation for Resource-Limited Devices

Deep Learning has shown superior performance on various complex tasks, such as object recognition and classification in the recent years. However, having said that methods based on deep learning have achieved near perfect accuracy, most of them are manually designed by the people who are expertise in the target field. In other words, building a good model is a laborious and time consuming task. But, today, the demand for machine learning boosted applications is constantly increasing and requiring more people to build them. Depth map estimation is an important task of computer vision, and is applied on several essential tasks, like the self-driving task. However, deep learning depth estimation models have performed promising results, they all usually need high computational devices to operate. Thus, it is urgent to solve this problem and propose a model for resource-constrained devices, such as drones, mobile devices, etc. Neural Architecture Search is a new point of discussion in the communities of industry and academics. The importance of Neural Architecture Search is it enables the automation of the designing process of neural architectures without the knowledge of the field. In this project, me and my supervisor have attempted to study Neural Architecture Search and apply it on the depth estimation task to create a novel architecture design. Keywords: Computer Vision; Deep Learning; Supervised Depth-Estimation; Neural Architecture Search; ProxylessNAS