Blurry Textures
Having some issues getting good textures in Reality Capture.
I know the dataset is decent because I can run it through other software and get really crisp detail. When I run it though RC it seems blurry. I feel like it has to be a setting somewhere that can improve this. Any suggestions?
*Just to note I am viewing both models in a the same 3rd part 3D viewer so they are fair comparison.
-
Is there a specific reason you chose Photo-consistency based?
The default value for this parameter is visiblity based. Perhaps changing this parameter will yield better results.
Also, maybe someone else can shed light on this but generating the texture on a 16x16k texture file should yield sharper results?
-
Hi Nick,
nice screenshooting, very easy to get all neccessary details! :-)
I'm afraid there is no magic button.
You could try and lower the texel size and see if that helps anything.
But to be honest I don't think it will. RC provides more detail in the mesh but it also requires "better" input. My best bet is that your alignment is slightly off, which is a common cause for blurry texture. Afaik RC has a different way of creating the texture. Rather than cutting one triangle from one image, it creates each pixel from an average of all suitable cameras. And if those are slightly misaligned, then it is bound to be blurry.
Jasper, with fixed texel size, there won't be any gain with a 16k texture map because RC will just use as many as it needs to fit all the pixels in there.
-
Maybe try to lower your max reprojection error until the components drops some images - try to find a sweet spot between 0.4 and 1 - 2 is maybe a bit too high.
Your alignment time is pretty fast (less than 3min in high for 922 pictures), what resolution are you using ? are you using RAW pictures ? (there is some Win drivers issues that kills the resolution of the images in the DNG driver...)
-
So I managed to cycle through a few different settings.
Firstly I performed a new alignment with 0.1 reprojection error. Anything more than that didn't drop off any additional photos.
Visability Based appears to give better results than Photo based.
Then between linear or multi it seems linear gave the best result.
Overall an improvement on what I had before however compared to the other results its still a few steps behind.
I'm wondering if the other package does some sharpening to the texture afterwards? I threw in the RC screenshot into Lightroom and cranked up the sharpness. Seems to produce a similar result.
-
Well, that was worth the effort! Thank you for sharing this!
Now with a bit of tweaking you get even better results than your other package!
I don't think that 0.1 is actually neccessary. DEVs say 1 should be enough for most cases. I tried with 0.5 once and to be honest I didn't see any difference any more. Even from 2 to 1 I didn't see much. I suspect that with 0.1 there will be a lot more images missing in the full image set, since mith more complexity more erros will be introduced. Which could be also the case why in your test it worked better - less complex, less errors...
Please sign in to leave a comment.
Comments
6 comments