Unsupervised Deep Homography A Fast and Robust Homography

Unsupervised Deep Homography: A Fast and Robust Homography Estimation Model Ty Nguyen, et al. GRASP Lab, University of Pennsylvania IEEE ROBOTICS AND AUTOMATION LETTERS, 2018 1

4

Supervised Approach • 6

Unsupervised Approach • 7

Unsupervised Approach • x • X 8

9

12

Experiment-Sythetic Data • MS-COCO dataset • Standardization: based on the mean and variance of pixel intensities of all images in our training dataset • Data Augmentation: inject random color, brightness and gamma shifts during the training 13

Experiment-Aerial Data • Standardization & Data Augmentation • Test sample: manually labeled the ground truth by picking 4 pairs of correspondences 14

15

Conclusion • Train a deep neural network to estimate planar homographies • Able to handle large displacements and large illumination variations • Better performance on the aerial image dataset than supervised approach • Future work • investigate robustness against occlusion • Unsuitable for change detection, try to use global information 16

Thanks! 17