Match-R-CNN-Repoduction

Reproducing the fashion image categorization and retrieval baseline approach from https://github.com/switchablenorms/DeepFashion2 / https://arxiv.org/pdf/1901.07973.pdf

Installation

conda install --file requirements.txt

Procedure

Data-preparation

download deepfashion2 dataset
run src/data/split_dataset.py on train dataset
run src/data/deepfashion2_to_coco.py with train, validation and test dataset

Train feature-network

run src/models/train_feature_network.py with train dataset

Feature and pair creation

run src/features/get_features.py with train, validation and test dataset
run src/data/create_all_pairs.py with train, validation and test dataset
run src/make_item_level_pairs.py with train, validation and test dataset

Train matching-network

run src/models/train_model.py