[Help] Surface noise and detail loss on High Detail reconstruction - RealityScan (Win 2.0)
Hi everyone,
I’m reaching out for some advice on improving mesh fidelity in RealityScan (Windows v2.0). I’m currently struggling with significant surface noise (bumpy/lumpy textures) and a loss of fine details in my final mesh, and I can’t seem to figure out the cause.
Here is the context of the project:
I am scanning a stone sculpture with physical dimensions of approximately 3m x 1m x 1.5m. My dataset consists of about 600 images, captured using a mix of a DSLR/Mirrorless camera and a smartphone.
My Workflow:
Pre-processing: I performed basic adjustments on the source images (corrected lens distortion and balanced exposure).
Alignment: The image alignment process seemed to go smoothly.
Reconstruction: I selected the “High Detail” setting to build the mesh.
The Issue:
Despite the high image count and pre-processing, the resulting mesh has a very uneven surface with a lot of “micro-bumps” or artifacts, and the sharp details of the sculpture are washed out.
I’ve attached some screenshots of the mesh and a sample source image for reference.
The biggest issue is that the final mesh lacks definition and fine details, which is confusing because the source photos are sharp and I believe my close-up coverage was adequate. However, I have two concerns regarding the dataset: some images have shallow depth of field (blurry backgrounds), and there is a slight white balance mismatch between the phone and camera shots. Could these inconsistencies be causing the surface noise? I’m unsure if I need to apply masks to the out-of-focus areas, strictly color-match the images beforehand, or if there are specific reconstruction parameters I should be tweaking. Has anyone encountered this problem before? I’m happy to provide more details if needed.
Hello Ahnfe,
such object can be tricky. It looks like it is all white, so it can be without important features for better alignment. It looks like there are some misalignemnts (according to the mesh result) Probably you will need to create some noise over the object (like by using the data projector, to get better mesh. Like there, but with other feature source: Tutorial: Using Scanning Spray to Scan Shiny Objects ). .
As you used also mobile, have you kept the camera priors active for such images?
Also, I woudl say that 600 images for such object is not enough. Also, the capturing path seems to be insufficient for me. You should cover the whole statue following the shape of it. As I can see on your screenshots the coverage is not so regular.
Thank you so much for the detailed explanation. You are correct; the subject is indeed a nearly pure white stone, which I understand is problematic for alignment.
To address your questions and clarify my constraints:
1、Accessibility: The sculpture is a fixed object, and I was physically unable to access the area behind it. Therefore, I could not capture the rear side. However, the back of the sculpture is uncarved/flat, so missing geometry in that specific area is acceptable to me.
2、Cannot Reshoot: This dataset was captured during a trip. Since it is an outdoor religious sculpture and I have already left the location, I cannot return to apply any physical treatments (like the scanning spray or pattern projection you suggested) or capture additional images. I am unfortunately limited to the current dataset.
3、Camera Priors: I am not entirely sure if the images contain usable Camera Priors (GPS/Pose data). I suspect the metadata might be incomplete. Since my operating system is in Chinese, I have attached a screenshot of the image file properties along with a translated version for your reference.
4、Software Solutions: Given that I cannot increase the image count or improve the physical capture conditions (as mentioned in point #2), are there any specific software techniques or reconstruction parameters I can tweak to squeeze better quality out of this existing data? Or is this simply the limit of the current dataset?
5、Decimation: To answer your final question—Yes, I decimated the model to 3 million triangles (3M tris) directly from the “High Detail” mesh.
Regarding the camera prior, you can see that in RealityScan. If there is a pin icon next to the images, then the priors were used. You can try to turn them off before the alignment.
To get better results you can tweak some alignment settings (like use a different distortion model with more parameters or set Image overlap to Small). You can also try to let the program find more features during the alignment.
But as you can’t recapture the object there is no a lot of things to try, as the capturing wasn’t ideal in my opinion.
Also, we recommend to simplify in 50% incremental steps down to the wanted triangle count.
