Master's thesis

Supervision by: Iro Laina, Dr. Federico Tombari
Thesis by: Andrei Militaru

Abstract

In recent years, deep learning based methods have been developed for accurate depth estimation of a scene, even from a single image. One limitation of these methods is that they require large data sets for training and, regardless of the generalization quality within the same distribution, previously unseen data from a new environment can still pose a challenge for a pre-trained model. Additionally, the data used for training is usually obtained by standard 3D cameras, like Kinect or ASUS Xtion. These devices have a limited range and can effectively be used for scanning a room, but they would not provide accurate results when scanning a long hallway or a building. The goal of this thesis is two-fold. First, we train a deep learning model for estimating the depth map from a single image, using a proprietary database of RGB images and their corresponding depth information acquired with a long-range 3D sensor. Second, we aim to improve the generalization capability of the model such that it adapts to depth distributions it did not encounter during training.