[학습]
(1) HRNet 기반의 pre-trained pose_hrnet 모델을 사용하였습니다. 
(2) augmentation:, flip, affine transform 등 수행
(3) Optimizer: Adam / Learning rate: 1e-4

[앙상블, 후처리]
(1) 앙상블 할 csv를 폴더에 넣어준 후,  ensemble_post_processing/ensemble_from_submit_csv.py 를 실행합니다.  (실행 시 경로 입력)
(2) 앙상블 후 ensemble_post_processing/postprocessing_from_csv.py 실행

[requirements]
(1)  code/requirements_hrnet_detection.txt 파일을 올려놓았습니다.
(2) detectron2  설치
- cuda 10.1, python 3.6.9, torch 1.4.0 환경에서 설치 후 첨부한 코드의 detectron2를 사용하였습니다.
- 직접 설치하실 경우 detectron2-0.2.1 버전을 설치하셔야 합니다.  

** 코드에서 final.ipynb'(학습데이터 1) / 'final_735.ipynb'(학습데이터 2) 파일을 실행하시고, 앙상블, 후처리 진행하시면 됩니다.

** 구글 드라이브 링크(코드, pre-trained 모델, 앙상블.후처리 등): 
https://drive.google.com/drive/folders/1iaq24HYD095dIo4O016vEC8f_jSs9-rI?usp=sharing

** 학습 이미지 위치
(1)  train_735.zip을 압축 푸셔서 이미지들을  final_sumission/images/train_735/images/train
(2) 전체 원본 학습이미지 받으셔서 final_sumission/images/train_imgs_origin/images/train 
(2) test 이미지는  final_sumission/images/test_imgs_origin