Skip to main content

Depth Estimation Example

Depth-estimation models estimate the relative distance from the camera for every pixel in an image.

Depth Estimation Model​

X-AnyLabeling includes Depth Anything V1 and Depth Anything V2.

  • Depth Anything V1 is a highly practical solution for robust monocular depth estimation by training on a combination of 1.5M labeled images and 62M+ unlabeled images.
  • Depth Anything V2 significantly outperforms its predecessor, V1, in terms of fine-grained detail and robustness. In comparison to SD-based models, V2 boasts faster inference speed, a reduced number of parameters, and enhanced depth accuracy.

Usage​

  1. Import images (Ctrl+I) or a video (Ctrl+O) into X-AnyLabeling.
  2. Select and load the Depth-Anything related model, or choose from other available depth estimation models.
  3. Click Run (i) to process the current image. After checking the result, press Ctrl+B to run the model on all images.

The output, once completed, will be automatically stored in a x-anylabeling-depth subdirectory within the same folder as your original image.

painting
Source
depth-anything-v1-gray
Depth Anything V1 (Gray)
depth-anything-v2-color
Depth Anything V2 (Color)
Tip

Two output modes are supported: grayscale and color. You can switch between these modes by modifying the render_mode parameter in the respective configuration file.

Advanced: Mapping Relative Depth to a Custom Range​

By default, these monocular models output relative depth: normalized values that indicate which regions are closer or farther. You can linearly map those values to a custom numeric range by adding the following parameters to the model configuration:

min_depth: 0.5       # Lower bound of the mapped range
max_depth: 20.0      # Upper bound of the mapped range
save_raw_depth: true # Save the mapped values as a .npy file

Example Configuration:

type: depth_anything_v2
name: depth_anything_v2_vit_b
display_name: Depth-Anything-V2-Base
model_path: depth_anything_v2_vitb.onnx
render_mode: color
min_depth: 1.0
max_depth: 50.0
save_raw_depth: true

When enabled, the output will include:

  • Visualization image: Color or grayscale heatmap (same as before)
  • *_depth.npy file: Depth values linearly mapped to the configured range

You can load and query the calibrated depth data using:

import numpy as np
depth_map = np.load("image_depth.npy")
value = depth_map[y, x]  # Get the mapped value at pixel (x, y)
Note

This operation is a linear remapping of relative model output; it does not turn a relative-depth model into a metric-depth model. Do not interpret the saved values as measured distances unless you have independently calibrated the model and camera for your scene. Leave min_depth and max_depth unset to keep the default visualization-only behavior.