Estimate Image Quality - inconsistent results

[Marcel]

We are struggling with picking out photos with motion blur using the Estimate Image Quality function. Sometimes images with higher than average estimated image quality are blurred anyway. At other times, images with low estimated quality are actually pretty sharp.

I did some tests with an artificial noise pattern, and the results are a bit inconsistent. When I add more blur to a noise pattern, the estimated quality actually goes up. The sharpest version has a lower estimated images quality than a version with a large amount of blur. Please see attached image with crops of the noise patterns and Image Quality values.

It would be great if the Estimate Image Quality fucntion would reliably pick out blurred images, because having to go through a sets of 400+ photos manually is no fun

 

[Alexey Pasumansky]

Hello marcel,

I think this is related to very small pattern structure (1 pixel?) and since PhotoScan estimates the quality by the border sharpness, here each pixel has its own border.

It will be interesting to see similar tests for bigger pattern size.

[Marcel]

Hi Alex,

You were right, the artificial pattern was the source of the inconsistent results.

But even with a real photo, the differences in estimated image quality are often still too small to differentiate properly. In the attached example you can see that the difference between a perfectly sharp and blurred image is only a value of "0.03" This value is so small that we cannot easily pick out the blurred images from a large batch of photos by sorting on image quality.

Perhaps it is because camera motion blur is often linear, so there are still a lot of sharp edges in the photo (perpendicular to the direction of the motion blur)?

[Alexey Pasumansky]

Hello Marcel,

I think I have forgotten to mention the most important thing about this feature (and I'll check if it is included in the Manual):
the image quality value refers to the area of highest quality. So, for example, images with low DOF will have high quality value, since focused area is sharp.

In such manner one can track bad images that do not have any sharpness and not good - without any blur.

[Marcel]

Yes, that makes a lot of sense when you have free standing objects with a blurry background.
But the estimated sharpness values still do not have enough separation to pick out bad images.

I've switched to running a Photoshop batch script that crops the center of the image, enlarges it by 200% (with nearest neighbour resize) and saves it as a different file. Then I flip through these images by hand and judge sharpness. It sucks having to go through 400+ files per project manually, but there is no other way to get perfect sharpness.

[Andrew]

Alexey, I only now noticed your explanation of Est. Image Quality referring to area of highest quality. This makes sense to help weed out photos with camera shake, but I think it would also make sense to have average IQ reported for entire photo, BUT, taking mask into consideration. A lot of times DOF issues are unavoidable and I end up taking a lot of photos, masking out different areas of them to only leave sharp areas. It would be great if Est. IQ feature could help determine whether enough out of focus pixels have been masked out.

Cheers,
Andrew