Agisoft Metashape

Agisoft Metashape => General => Topic started by: John_H on March 10, 2022, 01:36:11 PM

Title: Question about camera optimisation and RMS error statistics
Post by: John_H on March 10, 2022, 01:36:11 PM
Hi,

I'm using my normal workflow, which is to align the cameras, then improve the alignment using gradual selection. I will select points using
Gradual selection -> image count
Gradual Selection -> Reconstruction uncertainty
Gradual Selection -> Projection accuracy
and finally
Gradual Selection -> Reprojection error

At each stage I will delete a proportion of the 'worst' points, ensuring I keep the projections for each image above 150, and then optimise cameras.
At each stage I check the info to see how the process is affecting the RMS and Max reprojection errors. The latest project I ran had the following stats after the 1st, 3rd and final steps:

RMS reprojection error   0.190378 (0.648795 pix)
Max reprojection error   0.571035 (25.5418 pix)

RMS reprojection error   0.195382 (0.470916 pix)
Max reprojection error   0.745797 (3.44066 pix)

RMS reprojection error   0.134116 (0.318879 pix)
Max reprojection error   0.441291 (1.39499 pix)

My (first) question is, what do the two numbers mean - I assumed that they were the same value, just in different units, but the dimensionless number goes up while the pixel value goes down. What, then, is the first number, and which is the 'important' value (I assume it's the pix?)

Second, slightly related, is for the optimise camera options I have been ticking only:
f, cx, cy, k1, k2, k3, p1, p2
for the first three steps, and all the options for the final step. But the options in the newest version of Metashape have changed and I wondered if anyone had thoughts on which options to use now (especially the three new advanced options)?

Also, any thoughts and comments on the overall workflow would be greatly appreciated...

Title: Re: Question about camera optimisation and RMS error statistics
Post by: Paulo on March 10, 2022, 02:08:59 PM
Hi John,

the pix value are the real repro error at full 1x1 scale of imagery. So if you get a final RMS repro error of 0.31 pix then the rms reprojection for all your tie points is 0.31 pix * GSD. For a survey at 2 cm GSD then rms final is around 0.62 cm...
What happens is most points are actually sampled or determined at sub sampled imagery (2x2,3X3,4x4,,...) . So dimensionless number represents rms at sub sampled values. So basically you have following:

rms (pix) / rms (dimensionless) = Mean key point size or projection accuracy.

So it is normal that after 3rs step(Projection accuracy) the dimensionless rms varies less than pix rms.

For exampe for a final gradual selection on quality imagery, I get following numbers:
Code: [Select]
RMS reprojection error 0.0980884 (0.204929 pix)
Max reprojection error 0.317079 (0.759607 pix)
Mean key point size 2.07047 pix

So 0.204929 pix /  0.0980884 = 2.08922 pix which is about the same as Mean Key Point size 2.07047 pix....

So this means  that on average my keypoints were sampled or determined at imagery at 2x2 scale (i.e. for 4000 x 3000 image, keypoints were determined on 2000 x 1500 subsampled imagery)...(2.07)

Hope this makes it a little clearer,
Title: Re: Question about camera optimisation and RMS error statistics
Post by: John_H on March 10, 2022, 02:39:24 PM
Thanks Paulo.

GSD is ground survey distance? I am doing close range object photogrammetry, so the equivalent is just the size of each pixel on the object?

So I suppose my question is, in order to get the most accurate camera network, which number should I be concerned with? (ie, which one do I want to make smallest!). In the past I've always assumed that reducing the RMS pixel value was most important, but from your reply the mean key point size seems the most relevant value?

Thanks again for your help!

John
Title: Re: Question about camera optimisation and RMS error statistics
Post by: Paulo on March 10, 2022, 03:33:31 PM
Yes John,

for close range the GSD is equivalent size of pixel on object....

for good quality camera and good clear imagery, you should get RMS repro error (pixel) of 0.2 to 0.5 and mean key point size (which is projection accuracy) of between 2 and 3 pixel....after gradual selection....