PetFace (ECCV2024 Oral)

   

The official PyTorch implementation for the following paper:

PetFace: A Large-Scale Dataset and Benchmark for Animal Identification,
Risa Shionoda* and Kaede Shiohara* (*Equal contribution),
ECCV 2024 Oral

TL;DR: We established a large-scale animal identification dataset with more than 250k IDs across 13 families

⚠️Attention⚠️

Our PetFace dataset, code in this repository, and pretrained models are for non-commercial research purpose only.

Changelog

[2024/10/07] The code for face alignment on your own images was released.
[2024/09/02] Installation instruction was updated.
[2024/08/14] PetFace was selected as an ORAL presentation at ECCV2024 (with 3x Strong Accept!!!)🎉
[2024/07/27] Pretrained models, training code, and evaluation were released. Also, "split" folder is updated.
[2024/07/19] This repository was released.

Dataset

Fill in a google form for access to the dataset.

• We manually confirm the submitted forms and therefore it may take some time. Please be patient while the dataset link arrives.
• If you do not receive the link after a week, we may have missed your form or the form was not filled in correctly. We are sorry, but please submit the form again. (Please do not start a thread on Issue)
• We will not reply if there is missing information in the submitted form.
• We will not reply if you enter an e-mail address that is not affiliated with your organization (such as Gmail).
• We will not reply if the form is written in other languages than English.
• We will not reply if your company does not publish research papers.

Dataset directory

Place the dataset as follows:

. (Root of this repository)
└── data
    └── PetFace
        ├── images
        │   └── cat
        │       └── 000000
        │           └── 00.png
        ├── split
        │   └── cat
        │       ├── train.csv
        │       ├── val.txt
        │       ├── test.txt
        │       ├── reidentification.csv 
        │       └── verification.csv
        └── annotations
            └── cat.csv
         

train.csv: file names and id labels for training
val.txt: file names for validation (not used in this codebase)
test.txt: file names for verification (not used in this codebase)
verification.csv: pairs of file names to verify and labels indicating whether the pairs have the same ID
reidentification.csv: file names and id labels for re-identification

Setup

Docker (Recommended)

1. Pull a docker image:

docker pull pytorch/pytorch:1.12.0-cuda11.3-cudnn8-runtime

2. Replace the path in exec.sh.
3. Execute the dokcer image:

bash exec.sh

4. Install packages:

bash install.sh

pip (Unrecommended)

Install packages:

pip install -r requirements.txt

Testing

Pretrained weights are provided on google drive.

Re-identification

For example, you can run the evaluation of re-identification for cat as follows:

CUDA_VISIBLE_DEVICES=0 python3 src/reidentification.py -m arcface -w pretrained/arcface/cat.pt -i data/PetFace/split/cat/reidentification.csv -o results/reidentification/arcface/cat.csv

Then, you can compute the top-k (k={1,2,3,4,5}) accuracy:

python3 src/compute_topk_acc.py --topk 5 -i results/reidentification/arcface/cat.csv

Verification

For example, you can run the evaluation of re-identification for cat as follows:

CUDA_VISIBLE_DEVICES=0 python3 src/verification.py -w pretrained/arcface/cat.pt -i data/PetFace/split/cat/verification.csv -o results/verification/arcface/cat.csv

Then, you can compute AUC:

python3 src/compute_auc.py -i results/verification/arcface/cat.csv

Face Alignment on Your Own Images

We provide the source key points in keypoints folder to align images.
First, you need to detect 5 keypoints of your own image by AnyFace and save them as a .npy file.
Then, you can align the images by:

python3 src/face_align.py --tgt /path/to/your/keypoints.npy --img /path/to/your/image.jpeg --src /path/to/src/keypoints.npy --out /path/to/output/image.jpg

Training

For example, you can run the training for cat as follows:

CUDA_VISIBLE_DEVICES=0 python3 src/train_arcface.py src/configs/cat.py  --output outputs/cat/arcface

Also, you can train an arcface model on all the species (families):

CUDA_VISIBLE_DEVICES=0 python3 src/train_unified.py src/configs/unified.py  --output outputs/unified

Acknowledgement

We borrow some code from insightface, pytorch-center-loss, and triplet-loss-with-pytorch.

Citation

If you find our work useful for your research, please consider citing our paper:

@inproceedings{shinoda2025petface,
  title={PetFace: A large-scale dataset and benchmark for animal identification},
  author={Shinoda, Risa and Shiohara, Kaede},
  booktitle={European Conference on Computer Vision},
  pages={19--36},
  year={2025},
  organization={Springer}
}

Our Related Work

[GitHub] OpenAnimalTracks: A Dataset for Animal Track Recognition (ICIP2024)