Fine Grained Method for individual Identify

This repository contains code to reproduce the best results for this paper: zhirui, Animal Recognition Using Method of Fine-Grained Visual Analysis.

The best model achieved score one dev-hard-lost:

model	recall@1	recall@10	hit1_n90%@top0.1	hit10_n90%@top0.1
head_swin_bnneck	0.3229	0.6021	0.0	0.0282
head_swin_bnneck w/ trick	0.3177	0.6016	0.3474	0.2018

Reproduction of the best results

Install

Need python version 3.8 or later, linux OS, PyTorch==1.11.0 with cuda, install libararies list in requirements.txt.

git clone [email protected]:LostPetInitiative/study_spring_2022.git
cd study_spring_2022
cd zhirui
pip install -r requirements.txt

Download

From zenodo

# in zhirui folder
bash download.sh

From Kaggle

Go to kaggle dataset page, download all 4 file one by one.

1. unzip head_swin_bnneck.zip to ./download
2. unzip data_25.zip to ./download
3. unzip dev.zip to ./download
4. move yolov5s.pt to ./download

After that, the download folder shuold like this:

download/
├── data_25
├────── rf100199
├────── ...
├── dev
├────── found
├────── lost
├────── registry.csv
├── head_swin_bnneck
├────── cfg.yaml
├────── last.ckpt
├────── model.ckpt
├── placeholder
└── yolov5s.pt

Data Prepare

# in zhirui folder
# crop data_25 and dev images
# It will take some time to do it
python crop_image.py
# extract train/val data from data_25/body/head for model training
# It will take some time to do it
python train_data_split.py --data_dir ./download/data_25
python train_data_split.py --data_dir ./download/data_25_body
python train_data_split.py --data_dir ./download/data_25_head

Reproduc best result

Best model (head swin bnneck) trained on head crop dataset, so be care input the head crop dataset.

# in zhirui folder
python test.py --data_dir ./download/dev_head --save_dir ./download/reproduc --model ./download/head_swin_bnneck

Use threshold filter fake image for better hit metric.

# in zhirui folder
python test.py --data_dir ./download/dev_head --save_dir ./download/reproduc_trick --model ./download/head_swin_bnneck --filt 0.74

Train Model

Modify config file in config folder first.

# in zhirui folder
python train.py --config ./config/your_config_file.yml --device 0