Analysis of Unsupervised Loss Functions for Homography Estimation

Abstract

Neural networks proved their ability in complex classification and regression problems using labeled data. Recent trends have shown the impressive performance of neural networks in more complex problems like estimating ego-motion and homography tasks. Due to complexity and time consumption for labeling data, researchers tend to exhibit their attentiveness towards unsupervised data-based learning. However, there are no standard loss functions used for image reconstruction and less attention is drawn towards the loss functions than the end to end network architectures. In this paper, we carefully analyze and evaluate the two most commonly used loss functions for the homography estimation task.