4/23 :: Fast-up report

1. Abstract

• Goal of the Competition
  • Document Type Classification
  • Task : Image Classification
  • Evaluation Metric : Macro F1
• Timeline
  • Start Date : April 11, 2024
  • Final submission deadline : April 23, 2024 (19:00)
• Description of the work

  • Dataset overview

    • Class : 17 types of documents
    • Train : 1570
    • Test : 3140

    EDA

    • Analysis on label distribution of train data
    • Analysis on size/hue distribution of train and test data
    • Analysis on augmentation techniques used in test data

    Data Processing

    • Train data : data cleansing, upsampling, offline augmentation with augraphy
    • Test data : rectification
    • Online augmentation with Albumentation in models

2. Process : Competition Model

• Integrated version : using ResNet 152 and EfficientNet b4 to classify 17 types
• Individual version : using EfficientNet b4
  • step1 : 3-class classification of car/dashboard/documents
  • step2 : classification of 15 types of documents

    

    Individual version process

 

3. Process : Issues

• Impossible to find suitable rotation angles for every single test data
  
  • Build codes to detect angles by lines or object contour

4. Role

• EDA
  • Augmentation Analysis
• Preprocessing
  • conception & implemenation  for Test data
    • Rectificaiton 
    • Denoising
    • Test time augmentation
• Model Train
• Presentation

5. Results

• Public Score / Private Score
  • 0.9578 / 0.9441ㅇ
• Final standings of the Leaderboard

  

 

6. Conclusion

• 저번 대회에서 부족했다고 느꼈던 바가 있어 최대한 열심히 참여하려고 노력한 것이 도움이 된 것 같음
• 데이터 전처리를 위해 아이디어를 내고 구현해 적용하였고 모델 성능 향상에 기여해볼 수 있어 좋았음 
• 모델링 관련(fine, hyper-parameter tuning 등)에 좀 더 신경을 쓰면 좋았을 듯함