I would like to ask if removing blurred images is better than keeping them in the dataset.
Hi tw2016
Depending on the project - if you have enough overlap, then there is no need to remove them as the blurred images will not be used and only the overlapping/sharp images will be used…
But it is advisable to remove blurred images before starting to use them in a project…
Thanks.
Why is it advised to remove the blurred images when you say that they will (in the best case) not be used?
Sorry, I don’t understand this yet.
I think he is saying ‘if you have clear images that overlap, allowing you to reconstruct the scene, ditch the blurry ones. If ditching the blurry ones means you lose some of your hits in reconstruction, keep them in.’
Since my major projects start at 500 to 2k pictures, I keep the blurry ones in as taking them out sometimes means breaking the hit ratio on the over-all model (hard to tell which ones are gluing them together and it takes a while for RC to spit out each iteration). For smaller projects I run a model w/ blurred images and w/o them and use the one w/o them if it didn’t unglue everything.
Perfect reply, thank you!
I guess the developers themselves don’t know the answer. Must be a machine learning algorithm?!
Hi tw2016
The developers know it very well. And if you read if few times and test it on few datasets, you will see it in action.
So one more time :
If you have say 100 images looking at the same “spot” and say 95 are sharp and 5 of them are blurry, then the sharp images will outweigh the blurry ones in reconstruction and texturing ( overlap from sharp images in the blurry images – the same spot seen in sharp and blurry ones )…
Avoid blurry images if you can, but keep them in spots where you otherwise don’t have enough coverage.
Blurry images are a source of (sometimes much) less accurate alignment and bad/blurry texture since the algorithms have to guess a lot more where the features are.
Is there a way to find blurry images in RC like you can in Photoscan?
you can do that with photoscan? sounds interesting.
I’m sure its possible using opencv and the cli version.
but I don’t have any experience with opencv.
Yep. Select some or all images in “Photos” (not in “workspace”), rightklick and hit “estimate image quality”. It spews out 0.xy values and I think everything above 0.5 you might want to think about. It’s not 100% watertight, but a nice tool…
Götz Echtenacher wrote:
Yep. Select some or all images in “Photos” (not in “workspace”), rightklick and hit “estimate image quality”. It spews out 0.xy values and I think everything above 0.5 you might want to think about. It’s not 100% watertight, but a nice tool…
This is great, thanks!
One thing though… might want to think about keeping or might want to think about removing? I’m sure it’ll become apparent when I use it next round…
Hi all
This is a RealityCapture related forum !!!