Bilateral Filter: Should sigma_range produce identical blur in RGB vs CIELAB, or should the difference be documented?

[Ryan-Millard]

Hi everyone,

As noted in #192 and #191 something interesting about the bilateral filter occurs:

The filter can operate in either RGB or CIELAB color spaces, but using the same sigma_range produces different visual blur.

The difference

Original image

Filtered images (Left: CIELAB, Right: RGB)

Above:

• In the left image, the bilateral filter operated in the CIELAB color space.
• In the right image, the bilateral filter operated in the RGB color space.

For both images:

• pixels, width, height were identical - the same image was used.
• sigma_spatial was identical - 3.0 was used
• sigma_range was identical - 50.0 was used
• color_space was changed - the CIELAB image (left) used 0 and the RGB image (right) used 1

The RGB bilateral filter preserved the image the best in comparison.

Why this happens

• RGB channels: [0, 255] - max Euclidean distance ≈ 442
• CIELAB channels: L [0,100], a/b [-128,127] - max distance ≈ 374
• Real images: typical neighboring pixel differences in Lab are smaller than in RGB due to perceptual compression.

Result: Lab outputs look blurrier than RGB for the same sigma_range.

Experiment

I tested scaling RGB’s sigma_range to match Lab visually - a factor around 4.18 produces roughly the same blur for natural images.

Original image

Filtered images (Left: CIELAB, Right: RGB)

Above:

• In the left image, the bilateral filter operated in the CIELAB color space.
• In the right image, the bilateral filter operated in the RGB color space.

For both images:

• pixels, width, height were identical - the same image was used.
• sigma_spatial was identical - 3.0 was used
• sigma_range was identical - 50.0 was used
  • In the RGB color space, I scaled sigma_range by 1.18 - thus sigma_range was $4.18 \times 50 = 209$
• color_space was changed - the CIELAB image (left) used 0 and the RGB image (right) used 1

The RGB bilateral filter preserved the image the best in comparison.

Design question

For a library like this, what’s the better approach?

1. Automatically scale sigma_range internally so RGB and Lab produce visually similar results.
2. Leave sigma literal and document the difference, expecting users to control it themselves.
3. Optional: let users supply a custom scaling factor.

My thoughts

• Automatically scaling could confuse advanced users who expect the filter to behave according to the numeric sigma values.
• Leaving it unscaled is technically correct, but requires good documentation so users understand why RGB vs Lab outputs differ.

Conclusion

I’d love feedback from anyone who has implemented or used bilateral filters before:

• How do you usually handle this in libraries?
• Would you expect a library to match blur across color spaces automatically, or respect the numeric sigma values and document the difference?

Thanks in advance!

[Ryan-Millard]

Update

We'll document it and leave it up to the user.