Start

Install some libraries like torch, torchvision, e.t.c.

pip install -r requirements.txt

Data

Work with preparing datasets will be inside data folder. 200Gb for data and models should be enough.

cd data

Run below script to download MCS21 dataset (82GB) and unpack to data/train_data. May take 50 mins.

python prepare_mcs21_data.py

data/train_data/ should contain 2853 folders (each folder is unique video_id)

├── data/train_data/
│   ├── video_id_1
│   │   │── frame_1.jpg
│   │   │── frame_2.jpg
...
│   ├── video_id_2853
│   │   │── frame_1.jpg
│   │   │── frame_2.jpg

You may delete archive with data by rm -rf train_data.zip.

Now run below scripts, they will automatically download and preprocess a subset of OpenImages dataset and full 100DOH dataset. It may take 1-3 hours.

python prepare_oid6.py
python 100doh_convert_annotation2yolo.py

Your data folder should look like this:

├── data/
│   ├── train_data
│   ├── oid6_hand_yolo_100doh
...

There are should be 113425 images in data/oid6_hand_yolo_100doh/images/train, 2075 images in data/oid6_hand_yolo_100doh/images/val and 112009 annotation files in data/oid6_hand_yolo_100doh/labels/train

Hand detector

Hand detector training take at least 8 hours on 4x1080Ti

cd ../yolo_detector
python -m torch.distributed.launch --master_port 5216 --nproc_per_node 4 train.py --img 1024 --batch 24 --epochs 10 --data hand_oid6_100doh.yaml --weights yolov5m6.pt --device 0,1,2,3 --project 'runs/train/yolo5m6_1024_2x_100doh' --workers 16

You may take my trained detector google drive

Classication models

Train three classifiers. Training on 1x1080Ti will take at least 10 mins per model, 30 mins total.

cd ../classifier
python ./main.py --cfg ./config/r18_aug_crop0.7_final_antialias.yml
python ./main.py --cfg ./config/r18_aug_crop0.7_final.yml
python ./main.py --cfg ./config/r18_aug_crop0.6_final.yml

You may take my classifiers google drive

Now prepare submission

Copy trained yolo hand detector and classification models to ./submit folder

cd ..
rsync -avr ./yolo_detector/runs ./submit/ 
rsync -avr ./classifier/experiments/* ./submit/checkpoints/ 

Make an archive with submission

cd submit
zip -r mcs21_submit.zip *

Submit is now ready for uploading to boosters.pro