Warning: WORK IN PROGRESS. Currently the code is for demonstration purpose and does not run at the moment.
This project extends the deep homography estimation method from DeTone et al. [1] with a multi-scale strategy. Given a pair of input images, a homography is first estimated at the lowest resolution and then is progressively refined at higher resolutions. The training can be conducted using a synthetic dataset derived from the MS-COCO benchmark or other image/video datasets by following the method from DeTone et al [1].
The code builds upon Tensorflow(https://www.tensorflow.org/).
[1] DeTone, Daniel, Tomasz Malisiewicz, and Andrew Rabinovich. "Deep image homography estimation." arXiv preprint arXiv:1606.03798 (2016). https://arxiv.org/abs/1606.03798