idsia-robotics/micro-tartanair

# Mini-TartanAir: a 48x48 dataset for tiny Monocular Depth Estimation

## Dataset

This dataset scales down to 48x48 resolution 18 environments from the original TartanAir. In particular, we provide:

- tartanair_48x48: training split in 48x48 resolution (taken from the "easy" sequences)
- tartanair_hard_48x48: validation split in 48x48 resolution (taken from the "hard" sequences)

All images are annotated with 48x48 depth and disparity maps obtained directly from the synthetic ground truth. For every image, also a 360x360 centered portion of the original ground truth is provided, matching the same field of view of the respective image. If a 48x48 depth prediction related to a 48x48-sized image is upscaled to 360x360 resolution, these ground truth portions can be used to compute the prediction accuracy.

## Usage

To use the dataset, simply:
~~~
tar -xzvf tartanair_48x48.tar.gz -C ./
tar -xzvf tartanair_hard_48x48.tar.gz -C ./
~~~

Then, you can find the dataset under "tartanair_48x48/" (training) and "tartanair_hard_48x48" (validation).

## License

This dataset is derived from [TartanAir](https://theairlab.org/tartanair-dataset/), and therefore released under [Creative Commons Attribution 4.0 International License](https://creativecommons.org/licenses/by/4.0/) license.

## Citation

If you use this dataset, please cite:

~~~
@article{nadalini2025multi,
title={Multi-modal On-Device Learning for Monocular Depth Estimation on Ultra-low-power MCUs},
author={Nadalini, Davide and Rusci, Manuele and Cereda, Elia and Benini, Luca and Conti, Francesco and Palossi, Daniele},
journal={arXiv preprint arXiv:2512.00086},
year={2025}
}
~~~