Subscribe to download
Error's Kitchen

Nice capture!

RED-FROG- Thanks for the well thought out advice! It is very likely user error, though to be fair, quite frankly it's great to have something "just work." RealityCapture seems to do that by default for me.

I think you won't see the big differences in rendering speed until you add more photos (I saw it with 200). It's a scaling issue as far as I can tell.

On the steam page for RealityCapture they claim:

"The major competitive excellence: if you double inputs, the processing time will be doubled as well, but any other software will usually increase it fourfold."

I believe it based on what I've seen. I wouldn't have dreamed of using 1300 photos before but now I am and it's still taking a reasonable amount of time.

Hey.
Perhaps you would have to adjust tie and key point values in Photoscan to get better results. i.e. 60.00key and 20.000tie. (if you started lower than that)
Also make sure to disable "pair selection" when you get weird alignments and/or missing aligned photos. False aligned photos can be individually re-aligned in Photoscan.

I just had a quick test with RC and alignment time seemed pretty fast. 70 photos, ~1minute.
But the dense point cloud took a while. I think 18-20 minutes.

Photoscan took 10 minutes (vs ~1minute) for alignment but only 5 minutes (Medium) for the point cloud... (since the long alignment already resulted in a pretty dense Tie point cloud). So it was basically more or less the same speed?
I couldn't see anywhere how many points RC has generated and it was difficult to compare the clouds.. I just didn't see the advantage in using RealityCapture myself yet.
Guess I will have to play around more. I just need more time for it.

RED-FROG: Should add one more thing- in addition to getting many more holes on the same data with PhotoScan, I was also getting erroneous alignments and having things floating in the middle of where they shouldn't be in the renders. Not so with RealityCapture (again, on the exact same photos).

RED-FROG: Two things:
1. Processing speed, for sure. I can scale up a to a lot more photos because of this. This makes a big difference for me. I think they claim an O(n) algorithm vs O(n^2) on PhotoScan. Seems like it.
2. I've tried RealityCapture and PhotoScan on the exact same set of photos, and kept getting better results in terms of alignment and filling in holes. The kitchen for example- PhotoScan kept missing a good chunk of the oven and the fridge. RealityCapture had no problem with that (again, same set of photos).

It might be that RealityCapture is doing a better job of compensating for my less than perfect camera (GoPro).

Hello! What exactly do you mean by "works better"?
So far, Photoscan has always delivered good results and allows a lot of manual input. I haven't had the chance to test RC yet. But I'd like to do so, only to compare which one really works better.
RC does a lot of marketing about it's speed. But speed shouldn't cost quality of course.

Thanks tigs! Yeah it has definitely been trial and error for me (even with the resources online). One thing that would save me time knowing more than anything else- RealityCapture works a lot better than PhotoScan.

Nice work man.

I just discovered Destinations and doing this sort of thing is what I'm really keen on trying to nail on VR. Hopefuly can bounce some ideas back off each other!