Many times after Detect Markers, I have found out that the control point on some images are slightly off from the marker center, usually is at least half of out around 100 images, and it has lots of them have “red text pixels” (high error?). Does this affect the accuracy of the alignment? Do I need to correct those errors manually and delete the rest of the images for that marker? Since there are a lot of images, it won’t be ideal to correct all of them. Is there a quicker way to bulk delete images under a control point?
the optimal value for projection error is under 1 px, the highest acceptable should be around 2 px. The values above 3 px are showed with red triangle. These aren’t good measured and have bad influence for the alignment and the model precision. This errors depends on the point measurements but also on the quality of images and the way, how these images were taken.
You should correct these values or delete these very wrong measurements, if it is not possible to correct them. Unfortunately, you have to do it one by one.
374823146340, thank you for answering each of my questions, as I’ve asked multiple questions in a row. I actually forgot about this post and now I’m picking it up again.
So, judging from what you’ve said, I should correct or delete any yellow/red flagged images under a marker-control point, in order to get accurate marker positions, is that right?
It’s a bummer that we can’t bulk delete the images, when I shoot hundreds even thousands of photos, it will be time-consuming to correct/delete each one of the yellow/red-flagged ones.
Here comes my second question: also judging from what you’ve said, it seems that it’s not supposed to have so many errors over marker positions. How did that happen and how I can avoid them? Here’s my marker workflow: I place laser printer-printed paper markers on a flat surface around the subject and shoot with high resolution DSLR on a tripod. These are all done in a studio environment. So, I try to minimize the chance the markers got moved during the shoot and shoot with a very steady camera. The only things is that the markers in some of the shots will be out of focus, but RealityCapture will still detect them in those shots.
I thought with this setup and marker shots will be very good, but I still got yellow triangles on 1/5 and red triangles on 1/10 of the images of a marker. I tried to delete or correct some of them, but it usually will make some other unflagged images become yellow- or even red-flagged.
you don’t need to delete those images, just cancel measurement on them. How does look your alignment? Are there some images out of excepted position?
You can try to change the alignment settings, for example set different value in Max feature reprojection error.
Are those targets just laid on the ground? Or are they glued? If they are out of focus, this could also be the case of your problem. The point centrum is not so precious as on focused images.
You should delete all measurement, where these errors are bugger then 2 px.
Also, you can check your alignment with a Inspect tool, which could show you, which images are bad aligned or where you should take more images.
Yes, I meant cancel the calculations for control points, not actually exclude the images from the alignment.
I didn’t glue the markers as I shot them in a windless studio environment, but I shall try to glue them next time to completely eliminate the possibility of them moving.
For the inspect tool, it looks good to me, but I can only understand the camera relations, not the scene structure uncertainty lines. So I’ll just attach my inspect tool results for you to see.
I don’t get major error on the alignment, but minor ones, like when the areas that look like there is an extra layer of paint or something. I’ll attach a screenshot for this as well. I suspect those are caused by slightly misaligned markers. The image is showing what after I cancelled any images with 2 px & greater errors, so the situation is already improved but not been completely eliminated.
About the camera positions, it seems, that you have some weak parts here. According the Scene structure uncertainty (red and blue dots on the model), you have there these parts with red points. The areas with more red points are considered with higher uncertainty and therefore with lower precision. Add more images in such areas to improve the precision.
Also, pink color shows stronger connection between camera positions, blue is the weaker part.
About the misalignment, group your images w.r.t exif before running the alignment, if it is possible, you can use some tie point in the misaligned area. It might also help to align the images in several batches and keep adding images in smaller increments. If the resulting alignment from the smaller batch looks OK I would recommend to select all the images in this component and change Lock pose for continue to True. This will fix the relative positions among cameras in this component and during the next run connect more images to this existing one.