saicoco / kaggle-Steel-Defect-Detection

mxnet-gluon-based for Steel Defect Detection, unet achieves 0.904+

kaggle-Steel-Defect-Detection

mxnet-gluon-based for Steel Defect Detection with Kaggle-API, unet achieves 0.904+

Network

• UNet: resnet50
  • Data augumentation: filp up-down, left-right, random-crop .etc
  • Filter small instances
• Loss: Normalized softmax Focal loss refer to adaptis
• classification network: resnet50 + multi-classification

Usage:

You can use kaggle api to train or inference on kaggle platform, you can refer to kaggle-API