Algorithm

[LigasN]

I have started by loading input images and the first step after that was to convert them from sRGB (as blender returns them in that color space) to linear. That's the effect:

[LigasN]

Ok, that is wrong, I need to fix properties in the blender project.

[LigasN]

As in the camera's specification is written that possible outputs are:

• YUV 422/420 / YCbCr 422
• RGB565/555
• skompresowane dane 8-bitowe
• 8 / 10 bit RAW RGB Data

Then I would try with 8 / 10 bit RAW RGB Data, which I hope means that data is in linear color space.

[LigasN]

I've started to worry about algorithm ideas. For both light green N vectors, the difference in the captured image will be the same. I am talking about the brightness of a specific pixel. But I am leaving that wonder here for later 😄

[LigasN]

Slowly slowly 😄

[LigasN]

The angles idea with right conversion:

[LigasN]

And finally something after fix

[LigasN]

Expected similar to blender:

[LigasN]

Checked in Blender:

gif original link

[LigasN]

It is kinda bent to the down 😄

[LigasN]

with angles:

[LigasN]

Pixel position was wrongly calculated. Mistake with the Y sign. Fixed in this part of the commit. I am still debugging through the algorithm to ensure that everything is correct.

[LigasN]

No more issues found after that one, and normalmap now looks like:

[LigasN]

It is not perfect. It is deeper and somehow not balanced as for the small hemispheres on the gif from Blender below:

Normal map in blender

[LigasN]

Whats more, before I did't know about that color space, alpha etc. should be provided in Blender's settings for any input image from disk.

[LigasN]

And that is output from an algorithm that is using angles. I decided to proceed with the weight way to calculate normals from now on.