Structured Aleatoric Uncertainty In Human Pose Estimation

This provides codebase for the CVPR 2019 Workshop Paper

Note/TODO: Currently, only the evaluation code for pre-trained models and some skeleton code is provided. Yet to complete end-end training pipeline. This codebase and Readme.md build upon Integral Human Pose Regression codebase.

Loss Function

Network

Preparation for Training & Testing

1. Download MPII image from MPII Human Pose Dataset
2. Organize data like this

${PROJECT_ROOT}
 `-- data
     `-- mpii
        |-- images
        |-- annot
        |-- mpii_train_cache
        |-- mpii_valid_cache
     `-- hm36
        |-- images
        |-- annot
        |-- HM36_train_cache
        |-- HM36_validmin_cache

Usage

Test

To run evaluations on MPII Val dataset

Place the models in pytorch_projects/integral_human_pose/output/

cd pytorch_projects/integral_human_pose
python3 test.py --cfg experiments/hm36/resnet50v1_ft/d-mh_ps-256_dj_l1_adam_bs32-4gpus_x140-90-120/lr1e-3_u.yaml --dataroot ../../data/ --model output/covariance.pth.tar --is_cov True
python3 test.py --cfg experiments/hm36/resnet50v1_ft/d-mh_ps-256_dj_l1_adam_bs32-4gpus_x140-90-120/lr1e-3_u.yaml --dataroot ../../data/ --model output/diag.pth.tar --is_cov False

Cite

@article{gundavarapu2019structured,
  title={Structured Aleatoric Uncertainty in Human Pose Estimation.},
  author={Gundavarapu, Nitesh B and Srivastava, Divyansh and Mitra, Rahul and Sharma, Abhishek and Jain, Arjun},
  journal={CVPR Workshops},
  year={2019}
}