Workaround for low quality datasets (DOF, low coverage, ISO noise, etc) in RealityCapture


(Vlad) #1

Interesting workaround for low quality datasets (DOF, low coverage, ISO noise, etc) that RealityCapture for this moment can't successfully align.

You can align this images in Photoscan, disable preselection, and probably increase key/tie point limit. And using Gradual selection and Camera optimization optimize this alignment reprojection error to something small (with proper dataset, 0.1 px still possible).

And after this export Camera as Bandler *.Out and export undistorted images with {fileNUM}.{filext} file name mask (1.png,2.png,22.png,123.png,etc) in the same dir with bandle.out file. For data with masks probably better use PNG or TIFF.

Probably will be better remove all camera that not aligned. And for export use only aligned camera.
Or Bandle.out file can failed to open In RealityCapture (may be not required, but in case of troubles remember this)
And just open this bundle.out in RealityCapture and run mesh reconstruction. And later texture if needed.

Resulted mesh as usual have much more good details than just from Photoscan or even from ContextCapture. But I think using High resolution mode is not required because of low quality of original dataset. But you can try. At least it work good for me in High but also looks great in Normal.

Also you can try use this alignment as draft for RC and align again, but probably this will not increase quality too much.

This is example


small figurine i shot with iPhone with good overlap but with strong DOF on most surfaces.