FixMyPose / फिक्समाइपोज़
Code and dataset for AAAI 2021 paper "FixMyPose: Pose Correctional Describing and Retrieval" Hyounghun Kim*, Abhay Zala*, Graham Burri, Mohit Bansal.
Prerequisites
- Python 3.6
- PyTorch 1.4 or Up
- For others packages, please run this command.
pip install -r requirements.txt
Dataset
Please download resized images from here and unzip in dataset/fixmypose folder.
Also, you can download full-sized images from here.
Usage
To train the models:
# pose correction describing model (English)
bash run.sh
# pose correction describing model (Hindi)
bash run_hindi.sh
# target pose retrieval model (English)
bash run_retrieval.sh
# target pose retrieval model (Hindi)
bash run_retrieval_hindi.sh
Task Specific Metric Usage
python body_part_match.py ./path/to/transform_files/ [FILES...]
python direction_match.py [FILES...]
python object_match.py [FILES...]
[FILES...] - list of files you wish to run the metric on (e.g. output_val_seen.json output_val_unseen.json output_test_unseen.json ...). The files will be created in the root directory when you finish evaluations.
Hindi METEOR Calculation
cd langeval/cococaption/pycocoevalcap/meteor
python meteor_hindi.py hindi_meteor.json
If you uncomment these two lines when evaluating, you can obtain the hindi_meteor.json file in the root folder.
Evaluation on Test split
Please contact [email protected] for the test split.
Acknowledgments
Base code is from "Expressing Visual Relationships via Language" paper's code repository.
Caption evaluation code is from Microsoft COCO Caption Evaluation.