IDSIA-μMDE: a real world 48x48 dataset for tiny Monocular Depth Estimation
Dataset
IDSIA-μMDE is a real-world dataset designed for monocular depth estimation (MDE) under constrained IoT sensing conditions. The dataset was collected in an indoor laboratory environment using a nano-drone–based multi-modal sensing platform and is characterized by noisy depth annotations, reflecting realistic non-idealities of embedded perception systems.
The dataset contains RGB images acquired from multiple viewpoints of the same laboratory, featuring artificial lighting and various obstacles such as recycle bins, robots, and gates. Due to the sensing setup and platform constraints, depth labels are limited in range and feature a field-of-view mismatch with respect to the respective images, therefore representing noisy ground truth for both fine-tuning and evaluation.
IDSIA-μMDE is split as follows:
- Training set: 3,035 labeled images
- Validation set: 1,149 labeled images
- Test set: 3,171 labeled images
All splits contain RGB images paired with corresponding depth annotations.
Sensor specifications
Image data was collected with a OV5647 monocular camera (QVGA resolution, 320x240 pixels) with horizontal field of view of 57°. Images (sampled at 15 Hz) are downscaled to 48x48 RGB resolution with nearest neighbour interpolation from the QVGA resolution.
Depth ground truth was collected with the VL53L5CX Time-of-Flight (ToF) depth sensor, providing 8x8 depth maps with a range of [2 cm, 4 m] at 15 Hz and a field of view of 65°.
For further information about IDSIA-μMDE, please refer to the paper.
Usage
To use the dataset, simply:
tar -xzvf idsia-umde.tar.gz -C ./
Then, you can find the dataset under "idsia-umde/".
License
This dataset is released under Apache License 2.0 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}
}