sample volume measurement

[Mina]

Dear all,

I had three sediment boxes. I took a sample from each box. The idea was to measure the volume of sampling area in each box individually. For this purpose, the sediment surface was photographed with a camera with fixed focus before and after excavation.

The sediment surface in each box was simulated two times, before and after excavation and then the measured volumes were subtracted to calculate the sampling volume in each box.

All chunks were scaled using scale bar. Comparison of measured and simulated value for scale bars shows that the surface was accurately modeled in each chunk. You can find the project file here: https://1drv.ms/u/s!Ar3Xxh0xFee9xC9h8nXG0RSZu_LQ

I have still two questions to make sure that the model is working precisely:

1. Are the small scale-bar errors enough to conclude that each chunk was modeled accurately? or do I still need other criteria to check in terms of model accuracy?

2. Why the camera calibration parameters are different? Do I need to do optimization as well given that the scale bars errors are small right now?

Thank you for your feedbacks.

Mina

[n@sk]

Hi Mina,

I suppose that you used this imaging configuration because there is a practical rule of thumb on what you can expect in terms on horizontal and vertical accuracy but the fact that you are not using a pre-calibrated camera complicates things.
My (useless) answer is that it depends on the accuracy requirements.
In general, the model's accuracy could only be assessed if a reference model or target coordinates of x5-10 higher accuracy, which could be considered as ground truth, were available.
I do not know how exactly you referenced the chunks but, assuming that the before/after targets were not moved, you should use their coordinates to estimate the registration accuracy, which may be considered/assumed to be indicative of the overall accuracy.
Note, however, that there is no guarantee that they reflect what's happening in the middle of the model.
The scale bar errors are low but they can be significantly improved (at least 50% improvement for all chunks expect for the first one) if you optimise the alignment, even if you do not remove any potential outliers from the  sparse point cloud.
I do not see any reason why one would not click on a button that says "optimise" unless it is red or it is April fools' day


With regard to the calibration
the interior orientation parameter sets are different because  you selected the Adaptive Camera Model Fitting,
for example B2 was estimated for the last chunk (set this to zero and then optimise, unless you have reason to believe that it is essential in modelling your camera; I don't).
The value for each parameter is different in each chunk because their estimation is affected by random errors, just like the camera and point coordinates, and was based on different object points; in addition, because the images were not taken from the exact same positions and because the image planes are approximately parallel and coplanar.
If you had added convergent images or had accurate coordinates for some of the targets, I would say that you do not have to worry about it but if you plan on reporting height or volume differences you should think about how many decimal digits you will present.

I hope this helps.

[Mina]

Hi,

Thank you for your detailed response.

Regarding the targets, the before/ after targets were not moved during photography.  Your suggestion is to compare the estimated targets coordinates with each other in different chunks, right? What do you mean by “registration accuracy”?

In fact, since I don’t know how the optimization algorithm works and which camera parameters should be checked or not (like what you mentioned, B2 in the last chunk), I didn’t use it during my simulation. In addition, sometimes there is no improvement and I cannot interpret it exactly why (as you mentioned, no improvement happened for the first chunk). In general, is it possible to interpret somehow the adjusted camera calibration parameters to assess the quality of alignment step?

Regarding you last point, if I change the photography method to take convergent images, the results would be improved? Is there any problem with my photography method (parallel and coplanar)?

Thanks again for your explanation.

Mina

[n@sk]

I mean that if you use 3 targets to align each before/after chunk  pair, and the coordinates of the rest of the targets agree,  you will have an estimation of the internal accuracy of the model. If they do not agree then you cannot expect the external accuracy to be any better.
This is a more meaningful way of presenting the quality of the result in terms of 3D coords instead of the average  scalebar error.

You do not know how the optimisation works but it was used during the alignment phase anyway; you now need to fine tune it.
There are some points that are flying around, which I do not suppose that correspond to any object, thus they are false matches and should not be taken into account.
I can only see the low resolution thumbnails of the images but the reflections of the lights on the floor are not helping at all...

You do not know which parameters should be checked but some of them were checked and they were initialised to zero;
you should not leave this to chance and at least  try to be consistent.
Even if the values change, the same parameter set should be used  as you are using the same camera.
Since you are fixing the focus why don't you calibrate the camera so that you have a better initial guess for the parameters values and use it for all chunks?
Normally f,cx,cy,k123,p12 should be fine for a digital camera.

There's nothing you can deduce  from the calibration values about the metric accuracy of the 3d model.

The normal case is well studied, convenient when the surface can be assumed to be planar, in which case the accuracy can be assumed to be homogeneous and the images can be easily orthorectified, but is the worst choice with respect to  the uncertainty in depth estimation, especially with an uncalibrated camera.
I would also suggest adding images with varying roll angles.
The results should be improved but you shouldn't take my word for it.
Try measuring a planar surface or a sphere and see for yourself.

Having said all that, it depends on the accuracy requirements of your project; if this was gold you were measuring I would not buy from you...or maybe I would
Perhaps you already have what you need but judging from your original post you are aware of the fact that a third party might not be convinced, so you should at least be aware of all the caveats.

[Mina]

Hi

Thanks again for your suggestions.

Given that I didn’t have any 3D coordinates of targets (I couldn’t measure them indoors), scale bars errors are probably the only option to assess the accuracy of the model and the quality of the result. However, I agree that it cannot be enough and that was the reason for my first post to somehow improve the accuracy, if possible.

Regarding the camera calibration, once I tried to use Agisoft Lens, but I didn’t get consistent camera parameters for some trials, therefore, I preferred to trust the auto calibration results.

And for the last point, I am looking for around 1-2% of volume error which can increase significantly the accuracy of my further study.

Best,
Mina