I"m surprised that the AI upscaled worked as well as it did. As you know the sfm image matching process in photogrammetry depends on matching features between one image and another. With upscaled images, the AI generated features could be very different and may not be consistent across the images. In one image the AI could add one set of details to an image, and could add a very different set of details to another, thereby breaking or degrading the quality of the image matching. Thank you for doing this experiment!
fantastic... this is exactly what I was wondering myself so thank you very much for taking the time to do this test. I have a good understanding of the time it must have taken to do this... a lot of patience on your part for sure..... The only tiny criticism I have is when you was doing split frame 12:48native:48upscale. It would have been much easier to compare it it was all at the same time. not split screen but, 3 screens at the same time so it makes it easier to compare the frames... as in instead of splitting the frame into 3, take the same from from each, and lay them next to each other. If that makes sense?
I think it's a combination of the limitations of the lens, the relatively small sensor size, and the fact that it's a quad-bayer matrix. But mostly the size of the lens and sensor.
3:38 why doesn't it make sense ?! Half of 48mp is still better than 12mp... Moreover images from those small drones aren't exactly tack sharp and are highly manipulated by software already... For a DSLR I wouldn't say (depending on the quality of the lens) but for the small budget drones I don't think it makes a lot of sense to use ultra high profile...
Nice comparison video! try processing your raw files in DXO pure raw or Dxo photo lab. It is far superior than lighroom, also Metashape is not the best with texure generation.
I used Blender for the animation, but the Metashape animation is alright, although more bare-bones. Also, it is easier to align the different meshes in blender.
nice video, but you really wrong with quality parameters. In Metashape Ultra High preset means upscale image by factor of 4. So you basically used x16 upscaled images. Use High parameter instead.
I"m surprised that the AI upscaled worked as well as it did.
As you know the sfm image matching process in photogrammetry depends on matching features between one image and another. With upscaled images, the AI generated features could be very different and may not be consistent across the images. In one image the AI could add one set of details to an image, and could add a very different set of details to another, thereby breaking or degrading the quality of the image matching.
Thank you for doing this experiment!
Thank you very much for the time and effort you have invested in this video. Well done. Looking Forward to your next one.
fantastic... this is exactly what I was wondering myself so thank you very much for taking the time to do this test. I have a good understanding of the time it must have taken to do this... a lot of patience on your part for sure..... The only tiny criticism I have is when you was doing split frame 12:48native:48upscale. It would have been much easier to compare it it was all at the same time. not split screen but, 3 screens at the same time so it makes it easier to compare the frames... as in instead of splitting the frame into 3, take the same from from each, and lay them next to each other. If that makes sense?
Actually the 48 Mpx pictures doesn't have that much more details than the 12 Mpx pictures. For me the look almost the same
And dynamic range lost
An interesting experiment. Do you think that the 48mpx is already compromised by limitations of the lens?
I think it's a combination of the limitations of the lens, the relatively small sensor size, and the fact that it's a quad-bayer matrix. But mostly the size of the lens and sensor.
3:38 why doesn't it make sense ?! Half of 48mp is still better than 12mp... Moreover images from those small drones aren't exactly tack sharp and are highly manipulated by software already... For a DSLR I wouldn't say (depending on the quality of the lens) but for the small budget drones I don't think it makes a lot of sense to use ultra high profile...
Nice comparison video! try processing your raw files in DXO pure raw or Dxo photo lab. It is far superior than lighroom, also Metashape is not the best with texure generation.
I'll give it a try one day !
How did you do the animations? Agisoft or blender?
I used Blender for the animation, but the Metashape animation is alright, although more bare-bones. Also, it is easier to align the different meshes in blender.
5:13 There is hardly any difference. Maybe because how the video is rendered, but in the video i can't really tell.
Same accuracy different resolution. Just takes more time to process that´s all.
nice video, but you really wrong with quality parameters. In Metashape Ultra High preset means upscale image by factor of 4.
So you basically used x16 upscaled images.
Use High parameter instead.
no thats for the sparse cloud
read manual again. You are right.@@raek88