[Paper | Website | Colab (coming soon)]

Chenlin Meng, Enci Liu, Willie Neiswanger, Jiaming Song, Marshall Burke, David B. Lobell, Stefano Ermon

Stanford University

IS-Count is a sampling-based and learnable method for estimating the total object count in a region. It largely reduces the number of satellite images as well as human annotations compared to an exhaustive approach used by object detectors in many real-world counting tasks, while achieving a high accuracy.

• Overview
• Requirements
• Tutorials
• Data Preparation
• Running Identity Models
• Running Isotonic Models
• Citation

We consider the following four tasks across 45 countries:

1. Counting buildings in the US and 43 African countries
2. Counting cars in Kenya
3. Counting brick kilns in Bangladesh
4. Counting swimming pools in the US

The code has been tested on PyTorch 1.7.1 (CUDA 11.2).

To install necessary packages and dependencies, run

pip install -r requirements.txt
conda install gdal

The IS-Count pipeline could be divided into two steps: 1) data preparation and 2) object count estimation. We provide the code for the two steps with the example of estimating building count in New York State in the tutorials/ folder.

We provide code for preparing the necessary data for IS-Count in create_mask.py and create_data.py. Before running these scripts, you need to make sure you have the prerequisite data downloaded and organized as described in the file data/README.md. For more details on preparing the necessary data for running IS-Count, check out the data_prep_tutorial.ipynb notebook under data/.

To create the binary mask for the region of interest, run create_mask.py with the following command

python create_mask.py --sampling_method "$sampling_method" --district "$district" --overwrite

To create the all-pixel file for the region of interest, run create_data.py with the following command

python create_data.py --sampling_method "$sampling_method" --district "$district" --overwrite

We provide code to reproduce our results on counting buildings and brick kilns using uniform, NL-based, and Population-based identity models in baselines.py. To reproduce the results for all the three identity methods, run the following command:

for sampling_method in "uniform" "NL" "population"
do
  echo "$sampling_method"
  python baselines.py --sampling_method "$sampling_method" --percentage 0.02 --plot
done

We provide code to reproduce our results on counting buildings and brick kilns using uniform, NL-based, and Population-based isotonic models in isotonic_regression.py.

for sampling_method in "NL" "population"
do
  echo "$sampling_method"
  python isotonic_regression.py --sampling_method "$sampling_method" --percentage 0.0001 --plot
done

To get results for the Isotonic method in the paper, run with the --extra_train flag.

Please cite this article as follows, or use the BibTeX entry below.

@inproceedings{meng2022count,
  title={Is-count: Large-scale object counting from satellite images with covariate-based importance sampling},
  author={Meng, Chenlin and Liu, Enci and Neiswanger, Willie and Song, Jiaming and Burke, Marshall and Lobell, David and Ermon, Stefano},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={36},
  number={11},
  pages={12034--12042},
  year={2022}
}