breadcake / unsupervisedDeepHomography-pytorch

Pytorch implementation of Unsupervised Deep Homography.

Unsupervised Deep Homography - PyTorch Implementation

Unsupervised Deep Homography: A Fast and Robust Homography Estimation Model
Ty Nguyen, Steven W. Chen, Shreyas S. Shivakumar, Camillo J. Taylor, Vijay Kumar

cd code/

in code/ folder:

dataset.py: class SyntheticDataset(torch.utils.data.Dataset) implementation
homography_model.py: Unsupervised deep homography model implementation
homography_CNN_synthetic.py: Train and test

Preparing training dataset (synthetic)

Download MS-COCO 2014 dataset
Store Train and test set into RAW_DATA_PATH and TEST_RAW_DATA_PATH respectly.

Generate training dataset

It will take a few hours to generate 100.000 data samples.

python utils/gen_synthetic_data.py --mode train

Generate test dataset

python utils/gen_synthetic_data.py --mode test 

Train model with synthetic dataset

python homography_CNN_synthetic.py --mode train

Test model with synthetic dataset

Download pre-trained weights

链接：https://pan.baidu.com/s/102ilb5HJGydpeHtYelx_Xw  提取码：boq9 

Store the model to models/synthetic_models folder

python homography_CNN_synthetic.py --mode test

results

Release History

• 2021.4.5
  • Add TensorBoard visualization and some metrics.

Reference

https://github.com/tynguyen/unsupervisedDeepHomographyRAL2018