Predictive Inequity in Object Detection
- NVIDIA GPU (we used an NVIDIA V100)
- NVIDIA Drivers
- NVIDIA Docker
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:
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:
- We use the COCO API for performance evaluation.
- Run the evaluation (this will likely take over an hour).
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
- 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:
- Link the
datasetsdirectory as such:
ln -s ~/inequity-release/datasets ~/github/maskrcnn-benchmark/datasets
Edit the args in
augmented_loss_weightsis 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
5on individuals labeled as
DSin the classification network loss of Faster R-CNN (as described in the appendix of the paper).
To run training: