This is a deep convolutional neural network for estimating the relative homography between a pair of images. Deep Image Homography Estimation paper implementation in PyTorch.
- Backbone: MobileNetV2
- Dataset: MSCOCO 2014 training set
- Train/valid: generated 8,000/3,000 pairs of image patches sized 128x128(rho=32).
- Test: generated 5,000 pairs of image patches sized 256x256(rho=64).
- Python 3.7.3
- PyTorch 1.3.1
Extract training images:
$ python extract.py
$ python pre_process.py
$ python train.py --lr 0.005 --batch-size 64
If want to visualize during training, run in your terminal:
$ tensorboard --logdir runs
Unzip model file into .pt file. Homography Estimation Comparison on Warped MS-COCO 14 Test Set.
$ python export.py
$ python test.py
Method | Mean Average Corner Error (pixels) |
---|---|
HomographyNet | 3.53 |
SURF + RANSAC | 8.83 |
Identity Homography | 32.13 |