This is an unofficial TensorFlow implementation of Presence-Only Geographical Priors for Fine-Grained Image Classification
Prepare an environment with python=3.8, tensorflow=2.3.1.
Dependencies can be installed using the following command:
pip install -r requirements.txtPlease refer to the iNat 2018 Github page for additional dataset details and download links.
The original CNN predictions file used for evaluation can be downloaded from the official project website.
To train a geo prior model use the script train.py:
python train.py --train_data_json=PATH_TO_BE_CONFIGURED/train2018.json \
--train_location_info_json=PATH_TO_BE_CONFIGURED/train2018_locations.json \
--val_data_json=PATH_TO_BE_CONFIGURED/val2018.json \
--val_location_info_json=PATH_TO_BE_CONFIGURED/val2018_locations.json \
--model_dir=PATH_TO_BE_CONFIGURED/geo_prior_ckp/ \
--random_seed=42Other training hyperparams can also be passed as flags. For more parameter information, please refer to train.py.
To evaluate a model use the script eval.py:
python eval.py --test_data_json=PATH_TO_BE_CONFIGURED/val2018.json \
--test_location_info_json=PATH_TO_BE_CONFIGURED/val2018_locations.json \
--cnn_predictions_file=PATH_TO_BE_CONFIGURED/inat2018_val_preds_sparse.npz \
--ckpt_dir=PATH_TO_BE_CONFIGURED/geo_prior_ckp/| Prior | Classifier* | Dataset | Accuracy |
|---|---|---|---|
| No Prior [1] | InceptionV3 | iNat2018 | 60.20 |
| Geo Prior (no photographer) [1] | InceptionV3 | iNat2018 | 72.84 |
| Geo Prior (no photographer) (ours) | InceptionV3 | iNat2018 | 72.94 |
| Geo Prior (full) [1] | InceptionV3 | iNat2018 | 72.68 |
| Geo Prior (full) (ours) | InceptionV3 | iNat2018 | 72.84 |
*Classifier predictions are from the original paper [1].
[1] Original paper: https://arxiv.org/abs/1906.05272
[2] Official PyTorch code: https://github.com/macaodha/geo_prior
If you have any questions, feel free to contact Fagner Cunha (e-mail: fagner.cunha@icomp.ufam.edu.br) or Github issues.