2021 Ego-Vision 손동작 인식 AI 경진대회

Data Augmentation
- resize, blur, random brightness, random contrast, coarsedropout, ...

Optimizer & Scheduler & Else
- Adamp, CosineAnnealingLR
- batch size = 8

Models
- ResNet50, SeResNet50

Overview
- 중복된 이미지가 굉장히 많기 때문에, 최대한 heavy한 augmentation으로 배경에 치중되는 overfitting을 피하려고 했음
- 학습이 불안정한 만큼 배치사이즈와 LR이 성능에 큰영향을 끼쳤는데, 작은 배치와 적절한 LR조정을 통해 학습을 안정화 시켰음
- 단일 모델은 LB에서 0.10~0.11 정도를 달성하고 앙상블 모델이 조금 더 안 좋은 성능이 나왔지만 앙상블 모델을 최종으로 선택
- robust한 validation을 만들기가 어려웠기에, leaderboard의 성능에 의존하여 최종모델을 선택
- 추론속도 향상을 위해 폴더당 이미지 한장만으로 예측하였음

시도했던점
- keypoint 만 crop해봤지만, 성능에 큰 차이는 없었음
- postprocessing을 통해 class를 구분할까 생각했지만, logloss 특성상 logit값이 아닌 argmax값으로 예측하면 벌점이 굉장히 커지기에 private점수가 뒤바뀔까봐 시도하지 않음
- 다만 logloss trick은 사용해볼 여지가 있었고, 조금의 성능향상이 있었지만 깜빡하고 최종제출에서는 적용하지 못함
(참고 url : https://medium.com/@egor_vorobiev/how-to-improve-log-loss-score-kaggle-trick-3f95577839f1)

Our Github repo & Pretrained weights
- https://github.com/rlawjdghek/Dacon2021--Ego-Vision-