Predictive Inequity in Object Detection

Created by Benjamin Wilson, Judy Hoffman, Jamie Morgenstern at Georgia Tech

Please consider citing our paper if it's helpful to your research:

@article{wilson2019predictive,
  title={Predictive inequity in object detection},
  author={Wilson, Benjamin and Hoffman, Judy and Morgenstern, Jamie},
  journal={arXiv preprint arXiv:1902.11097},
  year={2019}
}

Prerequisites

• Docker
• NVIDIA GPU (we used an NVIDIA V100)
• NVIDIA Drivers
• NVIDIA Docker

Getting Started

Installation

• Install Docker: https://docs.docker.com/install/

• Install NVIDIA Docker: https://github.com/NVIDIA/nvidia-docker

• Clone the repository:

git clone https://github.com/benjaminrwilson/inequity-release.git ~/inequity-release/

• Pull the docker image:

docker pull benjaminrwilson/inequity-release:latest

• Run the docker image:

bash ~/inequity-release/docker/run.sh

Once you're within the container, you will need to get the necessary data to run the experiments listed in the paper. You will need to get the annotations (provided by us in MS COCO format), the images from the BDD100K dataset, and lastly the weights we used.

• Download the annotations, images, and weights:

bash ~/inequity-release/scripts/get_data.sh

Evaluation

• We use the COCO API for performance evaluation.
• Run the evaluation (this will likely take over an hour).

python ~/inequity-release/inequity/eval.py

The tables from the paper will be output as text based tables in a new folder called tables. The graph will be created in a folder called figs.

Training

• If you would like to train Faster R-CNN from ImageNet initialization, we have provided a training script to train at different weights. First, make a directory as such:

mkdir ~/weights/

• Link the datasets directory as such:

ln -s ~/inequity-release/datasets ~/github/maskrcnn-benchmark/datasets

• Edit the args in inequity/scripts/train_large.sh as needed. augmented_loss_weights is a list which corresponds to the weighting put on ["LS", "DS", "Not a Person", "A person, cannot determine skin type"]. For example, [1, 5, 1, 1] would put weight 5 on individuals labeled as DS in the classification network loss of Faster R-CNN (as described in the appendix of the paper).

• To run training:

bash ~/inequity-release/inequity/scripts/train_large.sh