[New Feature] Ensemble

[yoonghee]

What?

• 2 stage model & 1 stage model Ensemble

Why?

• for performance

How?

• 코드참고 : https://github.com/boostcampaitech2/object-detection-level2-cv-04/blob/main/output/ensemble.ipynb
• WBF weights tunning

Todo

• [x] SwinB, SwinB-rfp, SwinL, YOLOv5 ensemble
• [x] SwinB, SwinB-rfp, SwinL, YOLOv5, YOLOr ensemble
• [x] 다른 library ensemble
• [x] ensemble base code 추가

[yoonghee]

SwinB, SwinB-rfp, SwinL, YOLOv5 ensemble

WBF tunning

• SwinB - w = 1.1
• SwinB-rfp - w = 1.3
• SwinL - w = 1.3
• YOLOv5 - w = 1

Dashboard Score

• 0.6706

[yoonghee]

SwinB, SwinB-rfp, SwinL, YOLOv5, YOLOr ensemble

WBF tunning

• SwinB - w = 1.0714
• SwinB-rfp - w = 1.1428
• SwinL - w = 1.125
• YOLOv5 - w = 1
• YOLOr - w = 1.0357

Dashboard Score

• 0.6744