This project is developed for the CMP721 - Computational Photography (Fall 2022) PhD course given by Ahmet Selman Bozkır from Computer Engineering Department at Hacettepe University.
In this project, monocular depth estimation is implemented using pre-trained MiDaS deep learning model. Tensorflow Lite is used as a backend inference engine in the project. The engine uses the MiDaS v2.1 small as an inference model. The model can be downloaded from model's Tensorflow Hub page.
The model performance mostly depends on device specifications that run the application. The model tested on two devices detailed in table below:
| Device | OS | Inference Time (ms) | Frame Per Second |
|---|---|---|---|
| Xiaomi Redmi Note 10 Pro | Android 13 | 90-150 | 4 |
| Xiaomi Redmi Note 8 Pro | Android 11 | 70-80 | 6 |
| Lenovo K6 Note | Android 7 | 500-700 | 1 |
| Galaxy A02S | Android 12 | Failed | Failed |
| Galaxy J4 | Android 10 | 2400-2800 | 1 |
