Evaluation: UBC benchmark dataset
Data preparation is required to create dataset files in numpy format (.npy extension) for each sequence in UBC dataset. Use create_dataset_file.py from Deep Compare repository. The resulting data file must be renamed as [sequence_name]_data.npy.
To use the evaluation code for UBC benchmark dataset using pre-trained models, please install all these requirements.
eval_net.py
usage: python eval_net.py --data_dir=path_to_UBC_dataset --model_name=name_of_model --test_set='liberty' or 'yosemite' or 'notredame' --network_type='2ch' or '2ch2stream' or 'siam' or 'siam_l2'
extended_imagedatagen.py (a data generator required for batch generation)
Evaluation: HPatches benchmark dataset
The code is destined for descriptor computation on HPatches benchmark dataset. Follow the instructions to setup HPatches toolkit, data directories and additional requirements from https://github.com/hpatches/hpatches-benchmark. Siamese network trained on Liberty sequence is used by default. Descriptors are stored under ../data/descriptors/patch-match-net in HPatches descriptor format.
eval_net_hpatches.py
usage: python eval_net_hpatches.py
The code has been tested on Windows 7, 64-bit with Python 3.5.