Computer Vision 이상치 탐지 알고리즘 경진대회

불균형 데이터 셋을 학습하여 사물의 상태를 잘 분류할 수 있는 알고리즘 개발
Competition Link

• 주최 / 주관: Dacon
• Private 11th, Score 0.8830
• Final 9th (9/481, 2%)

---

Structure

Train/Test data folder and sample submission file must be placed under dataset folder.
If you want change dataset path, you can change in CONSTANT.py

repo
  |——dataset
        |——train
                |——10000.png
                |——....
        |——test
                |——20000.png
                |——....
        |——train_df.csv
        |——sample_submission.csv
  |——models
        |——model
        |——runners
  |——data
  |——utils

---

Development Environment

• Ubuntu 18.04.5
• i9-10900X
• RTX 3090 1EA
• CUDA 11.3

---

Install Dependencies (GPU)

sh install_dependency.sh

---

Solution

Train

• Fine-Tuned timm tf_efficientnet_b6
• Image Size (528x528)
• Focal Loss (alpha=0.25, gamma=5.0) with Label Smoothing (0.1)
• Trained for 70 epochs
• 5 StratifiedKFold train
• Train 30 epochs with mixup, trained remaining epochs without mixup
• Transpose, Resize, HorizontalFlip, VerticalFlip, ShiftScaleRotate(-30, 30), Normalize

python kfold_main.py

---

Inference

5 fold ensemble (soft-voting) with Test Time Augmentation

• HorizontalFlip, VerticalFlip

python kfold_inference.py

---

Tried Techniques

• CutMix
• ArcFace Loss
• Model Ensemble (EffNetB7 + EffNetB6)