ibaigorordo/midasv2_1_small-tflite-inference

Python scripts to perform monocular depth estimation using Python with the Midas v2.1 small Tensorflow Lite model.

Midas v2.1 small TFLite Inference

Python scripts to perform monocular depth estimation using Python with the Midas v2.1 small Tensorflow Lite model. Tested on Windows 10, Tensorflow 2.4.0 (Python 3.8).

Requirements

• OpenCV, Numpy and tflite (or tensorflow). pafy and youtube-dl are required for youtube video inference.

Installation

pip install numpy opencv-python tflite tensorflow
pip install pafy youtube-dl

Midas v2.1 small (link)

• Input: RGB image of size 256 x 256 pixels.
• Output: Inverse relative depth map with 256 x256 pixels.
• Inference speed: - 30 FPS on Iphone 11 NPU and 22 FPS on OnePlus8 GPU (Snapdragon 865).

Examples

• Image inference:

python imageDepthEstimation.py 

• Webcam inference:

python webcamDepthEstimation.py

• Video inference:

python videoDepthEstimation.py

Inference video Example

Original video: https://youtu.be/TGadVbd-C-E (by Nagasaki Biopark)

midasv2_1_small-tflite-inference's Issues

Working on Raspberry pi 4/5?

Have you had the chance to test the model on those devices? If yes, what FPS was achieved?

dislike count issue

since youtube removed dislikes there seems to be an issue when this part of the code is reached.

line 54, in _fetch_basic
self._dislikes = self._ydl_info['dislike_count']
KeyError: 'dislike_count'