Depth Estimation from a Single Omnidirectional Image using Domain Adaptation

Abstract

Omnidirectional cameras are becoming popular in various applications\nowing to their ability to capture the full surrounding scene in\nreal-time. However, depth estimation for an omnidirectional scene\nis more difficult than normal perspective images due to its different\nsystem properties and distortions. It is hard to use normal depth\nestimation methods such as stereo matching or RGB-D sensing. A\ndeep-learning-based single-shot depth estimation approach can be\na good solution, but it requires a large labelled dataset for training.\nThe 3D60 dataset, the largest omnidirectional dataset with depth\nlabels, is not applicable for general scene depth estimation because\nit covers very limited scenes. In order to overcome this limitation,\nwe propose a depth estimation architecture for a single omnidirectional\nimage using domain adaptation. The proposed architecture\ngets labelled source domain and unlabelled target domain data together\nas its input and estimated depth information of the target\ndomain using the Generative Adversarial Networks (GAN) based\nmethod. The proposed architecture shows >10% higher accuracy\nin depth estimation than traditional encoder-decoder models with\na limited labelled dataset.