video2depth

Depthmap video generator using machine learning

Supported Models

Model	Key	Description
MiDaS DPT-Large	midas_large	Default. Intel MiDaS large model
MiDaS DPT-Hybrid	midas_hybrid	Intel MiDaS hybrid model
MiDaS Small	midas_small	Intel MiDaS small (fast) model
Depth Anything V2	depth_anything_v2	Best quality/speed ratio
ZoeDepth	zoedepth	Metric depth (real-world distances)
Marigold	marigold	Diffusion-based, highest quality, slow

Sample Results

Original	MiDaS	Depth Anything V2	ZoeDepth	Marigold

Requirements

• Python 3
• FFmpeg

$ pip install -r requirements.txt
# Additional for Marigold
$ pip install diffusers>=0.25 accelerate

Usage

# Video to depth (default model: midas_large)
$ python video2depth.py --video [filename]
$ python video2depth.py -v [filename]
# Image to depth
$ python video2depth.py --image [filename]
$ python video2depth.py -i [filename]
# Specify model
$ python video2depth.py -v [filename] -m depth_anything_v2
$ python video2depth.py -i [filename] -m marigold

Testing

# Run all tests
$ python -m pytest test_video2depth.py -v
# Run a specific test class
$ python -m pytest test_video2depth.py::TestModelRegistry -v
# Run a single test
$ python -m pytest test_video2depth.py::TestVideo2Image::test_extracts_frames -v

Sample files in sample_image/ and sample_video/ are used by the tests.

Output

Results are saved to out/<input_filename>/:

File	Description
output_depth.mp4	Depthmap video
output_depth_sound.mp4	Depthmap video + original audio
output_merged.mp4	Original + depthmap (vertically stacked)
output_merged_sound.mp4	Original + depthmap + original audio