Trying to find the relative translation between 2 markers.

[Cayman]

And the fourth screenshot....

Screenshot04.jpg (81.89 KiB) Viewed 305 times

[andrew.kirillov]

Hello,

Hard to say for sure what is going wrong if you are sure that:
1) model definition/size is correct;
2) order of model points correspond to order of image points;
3) effective focal length is right.

If all the above is right, then we may need to go into details of algorithm implementation (see AForge.NET framework source code) and original papers explaining all the math of the algorithms.

P.S. Did you try to make left most point as the first one (as it was found by the previous poster)?

[Cayman]

Hi Andrew.
Yes, the four screenshots show the glyph with point #1 (yellow) in each of the four quadrants.
Projected mofel is same size as glyph.
Projected model points are defined in same order as glyph (top left, top right, bottom right, bottom left).
Focal length = inage width.

The pink outline is the recognized quadrilateral.
The white outlibe and cross shows the projected model using estimated pose info.
The colored axes show the axes of the white outline and white cross...not the pink one.
The white outline very closely matches pink when point#1 (yellow) is neither shallow nor deep in the image (i.e., yellow corner is pointed either left or right).
See the large inaccuracy apparent when point#1 (yellow) is either shallow (lowest) or deep (topmost)?

Basically, this test app lets me show the pink (quad) and white (estimated pose) simultaneously in real time....useful for understanding the behavior. And that leads to the other observation that inaccuracy does not increase and decrease smoothly, but rather in large "jerky" moves, suggesting maybe some rounding / precision issue or possibly something to do with smoothing.

I'll probaby delve into that paper after checking for any potentiol rounding, precision and smoothing issues.

Bedt Regards,
Pat