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

## Dataset

This dataset scales down to 48x48 resolution the [Eigen split](https://proceedings.neurips.cc/paper_files/paper/2014/hash/91c56ce4a249fae5419b90cba831e303-Abstract.html) of the KITTI dataset. It contains:

- A training split of 22600 images in 48x48 resolution
- A validation split of 888 images in 48x48 resolution
- A test split of 697 images in 48x48 resolution

All images are annotated with 48x48 depth and disparity maps obtained with the stereo-matching SGM algorithm. 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 kitti_48x48.tar.gz -C ./
~~~

Then, you can find the dataset under "kitti_48x48/". 

## License

This dataset is derived from [KITTI](https://www.cvlibs.net/datasets/kitti/), and therefore released under [CC-BY-NC-SA-3.0](https://creativecommons.org/licenses/by-nc-sa/3.0/) 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}
}
~~~