LCPN HierNet Mar 2021 - Jul 2021

Hierarchical image classification model for fashion commerce items based on EfficientNet-b4 and LCPN (Local Classifier per Parent Node) technique.

  • Hand-crafted a dataset of CH*NEL products using selenium (35 classes, 26693 images)

  • Implemented hierarchical image classification model for CH*NEL products in PyTorch.

  • Combined EfficientNet and Local Classifier per Parent Node technique to achieve 85.2% accuracy; 19.2% improvement from previous flat-classification model


Local classifier per parent node: training one multi-class classifier for each parent node, to distinguish between its child nodes.

© 2023 Tyler Kim. All Rights Reserved.