Agile Modeling

This is a companion codebase and dataset associated with the paper "Agile Modeling: From Concept to Classifier in Minutes".

Introduction

In our Agile Modeling paper, we describe a system that allows users without Machine Learning experience to create their own image classifiers from scratch, for any concept they have in their mind.

Here we release a Colab with step-by-step instructions guiding users through the Agile Modeling process.

Additionally, we also release all the data labeled by our users for the 14 concepts included in our experiments.

Setting up

Requirements

1. Python 3.7
2. TensorFlow 2.13.0
3. scikit-image

Part 1: Preparing the unlabeled data

While the Agile process requires no labeled data to begin with, we have to prepare the unlabeled pool of images from which the system selects a few images for the user to label.

In our experiments we used the "LAION-400-MILLION open dataset", but the Agile framework is not restricted to this. Any unlabeled data works as long as it is converted to the right format. We further describe how to download and preprocess the LAION dataset, but feel free to follow similar steps with your dataset of choice.

TODO: Add instructions on how to download and preprocess his data.

Part 2: Running the Colab

We will be releasing our code shortly.

Data collected during our user study

We will be releasing data shortly.

Citation

If you found this codebase useful, please consider citing our paper:

@inproceedings{agile_modeling,
  title={Agile Modeling: From Concept to Classifier in Minutes},
  author={Stretcu, Otilia and Vendrow, Edward and Hata, Kenji and
          Viswanathan, Krishnamurthy and Ferrari, Vittorio and Tavakkol, Sasan
          and Zhou, Wenlei and Avinash, Aditya and Luo, Enming and
          Alldrin, Neil Gordon and Bateni, MohammadHossein and Berger, Gabriel
          and Bunner, Andrew and Lu, Chun-Ta and Rey, Javier A and
          DeSalvo, Giulia and Krishna, Ranjay and Fuxman, Ariel
  },
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer
             Vision, {ICCV} 2023, Paris, France, October 2-6, 2023},
  year={2023}
}