This is an unofficial TensorFlow implementation of Presence-Only Geographical Priors for Fine-Grained Image Classification

Requirements

Prepare an environment with python=3.8, tensorflow=2.3.1.

Dependencies can be installed using the following command:

pip install -r requirements.txt

Data

Please 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.

Training

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=42

Other training hyperparams can also be passed as flags. For more parameter information, please refer to train.py.

Evaluation

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/

Results

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].

Source

[1] Original paper: https://arxiv.org/abs/1906.05272

[2] Official PyTorch code: https://github.com/macaodha/geo_prior

Contact

If you have any questions, feel free to contact Fagner Cunha (e-mail: [email protected]) or Github issues.

License

Apache License 2.0