Identification via Metric Learning

This repository contains the source code that accompanies our paper "Visual Identification of Individual Holstein Friesian Cattle via Deep Metric Learning" - available at https://arxiv.org/abs/2006.09205. At its core, the code in this repository is adapted and extended (with permission) from Lagunes-Fortiz, M. et al's work on "Learning Discriminative Embeddings for Object Recognition on-the-fly" published in ICRA 2019 - paper, source code.

Within our paper, the code in this repository relates to section 5 on the "Open-Set Individual Identification via Metric Learning" and the experiments conducted in section 6. A selective set of the highest-performing weights from the experiments on open-set identification are included in this repository at weights/.

Installation

Simply clone this repository to your desired local directory: git clone https://github.com/CWOA/MetricLearningIdentification.git and install any missing requirements via pip or conda: numpy, PyTorch, OpenCV, Pillow, tqdm, sklearn, seaborn. This repository requires python 3.6+

Usage

To replicate the results obtained in our paper, please download the OpenCows2020 dataset at: https://data.bris.ac.uk/data/dataset/10m32xl88x2b61zlkkgz3fml17. Place the contents (two folders: images/ and splits/) of the identification folder in datasets/OpenCows2020/. A selective set of weights from the paper are included in the weights folder.

To train the model, use python train.py -h to get help with setting command line arguments. A minimal example would be python train.py --out_path=output/ --folds_file=datasets/OpenSetCows2020/splits/10-90.json. To train on your own dataset, write your own dataset class for managing loading the data (similarly to datasets/OpenCows2020/), import it into utilities/utils.py and add the case to the def selectDataset(args) method.

To test a trained model by inferring embeddings and using KNN to classify them, use python test.py -h to get help with setting command line arguments. A minimal example would be python test.py --model_path=output/fold_0/best_model_state.pkl --folds_file=datasets/OpenSetCows2020/splits/10-90.json --save_path=output/fold_0/.

To visualise inferred embeddings using T-SNE, use python utilities/visualse_embeddings.py -h to get help with setting relevant command line arguments. A minimal example would be python utilities/visualise_embeddings.py --embeddings_file=output/fold_0/test_embeddings.npz

Citation

Consider citing ours and Miguel's works in your own research if this repository has been useful:

@article{andrew2020visual,
  title={Visual Identification of Individual Holstein Friesian Cattle via Deep Metric Learning},
  author={Andrew, William and Gao, Jing and Campbell, Neill and Dowsey, Andrew W and Burghardt, Tilo},
  journal={arXiv preprint arXiv:2006.09205},
  year={2020}
}

@inproceedings{lagunes2019learning,
  title={Learning discriminative embeddings for object recognition on-the-fly},
  author={Lagunes-Fortiz, Miguel and Damen, Dima and Mayol-Cuevas, Walterio},
  booktitle={2019 International Conference on Robotics and Automation (ICRA)},
  pages={2932--2938},
  year={2019},
  organization={IEEE}
}