This repository is the official implementation of:

Metadata Normalization
Mandy Lu, Qingyu Zhao, Jiequan Zhang, Kilian M. Pohl, Li Fei-Fei, Juan Carlos Niebles, Ehsan Adeli
CVPR 2021

The proposed Metadata Normalization operation. MDN layer takes the learned features from the previous layer (f), analyzes the effects of the metadata on them, residualizes such effects, and outputs the distribution corrected features (r).

The video presentation of this work at CVPR can be viewed here.

To install requirements, run:

pip install -r requirements.txt

with Python 3 (3.8 used).

synthetic_data_vis.ipynb provides step-by-step instructions for generating synthetic data and visualizes the inputs.

To train the model(s) in the paper, run this command:

python train.py --mdn <mdn_type> 
                --model_dir <params_and_output_path> 
                --batch_size <batch_size> 
                --epochs <num_epochs> 
                --N <group size> 
                --runs <num_runs> 
                --seed <random_seed> 
                --lr <learning_rate>

or use

python train.py

for the default mode.

If you use this code in your research, please cite our paper.

@inproceedings{lu2021metadata,
  title={Metadata Normalization},
  author={Lu, Mandy and Zhao, Qingyu and Zhang, Jiequan and Pohl, Kilian M and Fei-Fei, Li and Niebles, Juan Carlos and Adeli, Ehsan},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={10917--10927},
  year={2021}
}

These resources were used or cited within the code:

• Synthetic Data Generation paper code.

Mandy Lu, [email protected]