(Still under testing)

• An Pytorch re-implementation for a Text-to-image Re-ID algorithm
• This is an implementation of ICCV'19 paper Improving Description-based Person Re-identification by Multi-granularity Image-text Alignments. Refer to the original paper and authors for details.

@article{niu2020improving,
  title={Improving description-based person re-identification by multi-granularity image-text alignments},
  author={Niu, Kai and Huang, Yan and Ouyang, Wanli and Wang, Liang},
  journal={IEEE Transactions on Image Processing},
  volume={29},
  pages={5542--5556},
  year={2020},
  publisher={IEEE}
}

• This is a beta version implementation. Bug could exist.
• If you find this is useful in your research work, please cite the original paper (and probably star this repo ;-))

Please download the train and val1 setfrom WIDER Person Search by Language dataset and save it in proper folder.

With bi-GRU as caption encoder and ResNet-50 as image encoder, we got the following results:

for computational reason, we haven't implemented the global-global + global-part + part-part versiono of MIA, which might be released in next updates.

run sh src/run.sh

run sh src/run.sh with --mode val and --load_ckpt_fn set as the path to the saved checkpoints.

1. Check notebook src/inference.ipynb for interactive retrieval with config set properly

2. run python src/visualizations/heep_generator.py for generate a html page visualize retrieval result. Please set parameter accordingly to enable different visualization. e.g. False only for only display retrieval failures.