ePillID Benchmark

This repository contains data and code for ePillID - a benchmark for developing and evaluating computer vision models for pill identification. The ePillID benchmark is designed as a low-shot fine-grained benchmark, reflecting real-world challenges for developing image-based pill identification systems.

The characteristics of the ePillID benchmark include:

• Reference and consumer images: The reference images are taken with controlled lighting and backgrounds, and with professional equipment. The consumer images are taken with real-world settings including different lighting, backgrounds, and equipment. For most of the pills, one image per side (two images per pill type) is available from the NIH Pillbox dataset.

• Low-shot and fine-grained setting: 13k images representing 9804 appearance classes (two sides for 4902 pill types). For most of the appearance classes, there exists only one reference image, making it a challenging low-shot recognition setting.

• The best performing model at the time of publication is a multi-head metric learning approach. Following a CNN encoder, a bilinear transformation layer is used to represent fine-grained embeddings.

Please refer to our paper for more details:

@inproceedings{usuyama2020epillid,
  title={ePillID Dataset: A Low-Shot Fine-Grained Benchmark for Pill Identification},
  author={Usuyama, Naoto and Delgado, Natalia Larios and Hall, Amanda K and Lundin, Jessica},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops},
  year={2020}
}

Data

Downloadable from the releases page.

Docker

Docker image https://hub.docker.com/r/naotous/pytorch-image

sudo nvidia-docker pull naotous/pytorch-image:py36torch041-legacy

Train and evaluate models

sudo nvidia-docker run --ipc=host -it -v $HOME:$HOME naotous/pytorch-image:py36torch041-legacy /bin/bash

use -v mount option to access folders from Docker containers

source activate py36

python train_cv.py --appearance_network resnet18 --max_epochs 10 --data_root_dir /home/naotous/ePillID_data

NOTE: max_epochs 10 for quick testing

Train and evaluate models on Azure ML

python azureml_submit_runs.py ePillID_0430

Notes and acknowledgments

The ePillID dataset includes customized data from two public NIH NLM databases. Please refer to their websites for original data:

• NIH NLM Pill Image Recognition Challenge dataset
• NIH NLM Pillbox

For bilinear pooling implementations, we forked fast-MPN-Cov GitHub repo.

Please consider citing their work as well when you find our work useful in your research.

Disclaimer

This dataset and software are released for research purposes only.