Unofficial PyTorch implementation of GLAMpoints: Greedily Learned Accurate Match points

Unofficial PyTorch implementation of GLAMpoints: Greedily Learned Accurate Match points. The majority of code is based on the repository https://gitlab.com/retinai_sandro/glampoints

Requirements

Please, use Python 3, install PyTorch 1.4, OpenCV and additional libraries from requirements.txt

Datasets and Training

In order to re-train network please use PS-Dataset, train/test split is already prepared in datasets/ps_dataset/

Training configurations and paths to datasets are stored in configs/glampoints_training.yml.

python train.py --path_ymlfile configs/glampoints_training.yml

Logs and checkpoints are stored in tensorboard format in the directory logs/experiment_name/

Validation on HPatches-sequences

Validation code is adapted from D2-Net evaluation on HPatches

To run validation on ported version of weights please use

python evaluate_hpatches.py --path_hpatches - path_to_hpatches_sequences --init_weights init ---path_ymlfile glampoints_eval_ported_weights.yml --name glampoints_retina

To run validation on trained on PS-dataset version please use

python evaluate_hpatches.py --path_hpatches - path_to_hpatches_sequences --init_weights modified ---path_ymlfile glampoints_eval.yml --name glampoints_retina**

To create plots, please downlod results of other methods from D2Net repo and use Add in methods,

python eval/generate_hpatches_plot.py --path_to_hpatches_sequences --path_to_cache_dir

Ported weights from Tensorflow implementation

Trained network on PS-Dataset