This repositaty contains the code and trained models for our WACV 2019 paper: "Aligned to the Object, not to the Image: A Unified Pose-aligned Representation for Fine-grained Recognition". It contains three major parts: 1. pose estimation network, 2. patch feature extractors, and 3. classification network. The later two parts are in the process of final cleaning and will be available soon.
Note this repo is still being upadted so some inconsistency should exist. I'll try fix those ASAP.
We have two sub-modules for pose estimation frametwork. One is written in Torch (Lua) with ResNet-50 as back-bone network. The other is written in Pytorch (Python) with ResNet-34 as back-bone network.
Now the patch feature extractor network is added as Feature-Extraction folder.
Torch needs to be installed in order to train.
Specify your own dataset path and pretrained model path in simple.sh
.
Simply run the second line of simple.sh at Feature Extracion Folder. HDF5 should be installed.
TODOS: