※ 이 페이지에는 사용하는 주요 py파일에 대해서 간략히 서술만 했습니다. 실제로 구동을 원하신다면 아래 사이트의 Readme.md를 참조하여 코드를 사용하시면 됩니다.
https://github.com/Hwan-I/motion_keypoint 

모델 : HRNet
SOTA : https://paperswithcode.com/paper/deep-high-resolution-representation-learning
코드 : https://github.com/leoxiaobin/deep-high-resolution-net.pytorch#readme

참조 사이트 
잘못된 키포인트 : https://dacon.io/competitions/official/235701/talkboard/402730?page=1&dtype=recent
baseline : https://dacon.io/competitions/official/235701/codeshare/2411?page=1&dtype=recent

HRNet을 활용해 keypoint Detection과 bbox Detection 2가지를 수행했습니다. 우선 bbox model로 bbox를 만들고 해당 데이터와 keypoint, image 데이터를 기반으로 최종 keypoint를 예측했습니다.

기타
- loss 그래프를 봤을 때 lr_scheduler, learning_rate를 조절했으면 더 좋은 성적을 거두었을 것 같습니다. 
- 그 외 augmentation 완화, HRNet의 Hyperparameter 조정, pre-trained 가중치 일부 freeze 등을 변경하여 실험해도 좋을 것 같습니다.

추가 
Private Score 복원 코드를 Google Drive에도 게재했습니다. 아래 링크에서 다운받고 setup_guide.md를 따라 진행하시면 됩니다.
https://drive.google.com/drive/folders/1PypA1xBjMVkg7pTaWb9a7INMaY0MMDCH?usp=sharing