PkuRainBow/HDC.caffe

Implements a cascaded deep metric learning framework with hard example mining for fine-grained visual recognition tasks like product and car classification. The approach uses custom Caffe layers—NormalizationLayer and PairFastLossLayer—to enable hard-aware pair sampling during training, optimizing embedding spaces for retrieval benchmarks like Stanford Online Products, CARS196, and CUB-200. Extends Caffe with configurable loss parameters supporting positive/negative/both pair mining modes with learnable margin strategies.

No commits in the last 6 months.