depth_anything

Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data

Input

• Image or Video

The script will perform a monocular depth estimation on the input media.

Output

• Depth image

Estimated relative depth with inferno colormap(without option -g), or single channel grey scale image(with option -g).

Saves to ./output.png by default but it can be specified with the -s option

Usage

Internet connection is required when running the script for the first time, as it will download the necessary model files.

Running this script will estimate the relative depth of the input image/video. The results will be shown in a separate window(when inferencing on image and video), or saved as an image(when inferencing on image).

Example 1: Inference on prepared demo image.

$ python3 depth_anything.py

The result will be saved to output.png by default.

Example 2: Specify input path, save path, and encoder type.

$ python3 depth_anything.py -i input.png -s output.png -ec vitl

-i, -s, -ec options can be used to specify the input path, save path, and encoder type separately.

Example 3: Inference on Video.

$ python3 depth_anything.py -v 0

argument after the -v option can be the device id of the webcam, or the path to the input video.

Reference

• DepthAnythig

Framework

Pytorch

Model Format

ONNX opset=11

Netron