chohoseong / p3-ims-obd-code-chaser

p3-ims-obd-code-chaser created by GitHub Classroom

TEAM : Code Chaser

• member
  • 반영성(https://github.com/ys-ban)
  • 소재열
  • 이대훈
  • 조호성(https://github.com/chohoseong)
  • 최이서(https://github.com/iseochoi)

Task and solving strategies

• data: a part of TACO dataset
• problem and breakthrough:
  • overfitting -> multilabel stratified 5-fold cross validation, lr scheduling, weight decay
  • inconsistency of images -> augmentations(flip, grid mask, rotate, etc)
• semantic segmentation
  • architecture: UNet3+ with depp supervision and class guide module
  • backbone: EfficientNet-B3(pretrained, noisy student)
  • validation: multilabel stratified 5-fold cross validation
  • ensemble: 5-fold soft voting
• object detection
  • model1:
    • backbone: Swin Transformer
    • neck: FPN
    • detector: Hybrid Task Cascade
  • model2:
    • backbone: EfficientNet-B5
    • neck: BiFPN
    • detector: EfficientDet
  • model3:
    • backbone: CSP-Darknet (depth multiple: 1.33, width multiple: 1.25)
    • neck: PANet
    • detector: YOLO
  • multi head ensemble based on WBF

Result

• semantic segmentation
  • mIoU - 0.6795
• object detection
  • mAP@50 - 0.4171