CompFashion

A code base for conducting image retrieval with text feedback experiments.

Available methods:

• For visually compitable garment retrieval:

  • CSA-Net (CVPR2020)

• For text-guided garment retrieval:

  • TIRG (CVPR 2019)
  • VAL (CVPR 2020)
  • CoSMo (CVPR 2021)
  • RTIC w/o GCN (arXiv 2021)

Data Preparation

1. Download FashionIQ dataset according to this. Then move data into train_data/fashioniq/.

2. Build vocabulary.

   python -m nltk.downloader 'punkt'
   python build_vocab.py

3. (Optional) Generate GloVe embeddings.

   python -m spacy download en_vectors_web_lg
   python gen_glove.py

Running

python train_net.py --config-file configs/fashioniq/tirg_bigru_init.yaml --use-tensorboard

Results

coming soon...

Different Settings in Original Methods

1. VAL uses smaller val split and GloVe.
2. CoSMo constructs captions with both orders for concatenating.
3. ...