A few months back I was at the Natural History Museum, London, collecting photogrammetry data of dinosaur feet. I got plenty of data, but by the end of the day I’m ashamed to say I was rushing a little bit, and trying to get data off my camera, keep it charged, and get myself ready for a pint.
The result was that a couple of the phalanges of Iguanodon were collected… poorly. Well, see for yourself:
One of these was taken with my Sony nex-6. I got 17 photos like that, taken in my light box. But, no scale bar was present, and the 17 photos weren’t enough for either Meshroom or Metashape to fully reconstruct the bone.
I also had 42 images taken with my phone (note 8), but the scale bar moved around inconsistently between them. Unsurprisingly, I really struggled to get any photogrammetry software to reconstruct the bone.
I started trying to mask the images in Metashape, but the tools there are not particularly smooth or quick to use – it can take minutes per image. This was made worse because the small size of the bone meant a lot of the time it wasn’t entirely within the focus zone of the camera, so Metashape’s automated masking tools struggled to find edges.
Then I remembered reading about the new ‘select object’ tool in Photoshop 2020. I try and use free software when I can, but I get access to adobe software through work, so thought what the hell.
Anyway, long story short, the results are incredible – it’s insanely accurate, and insanely fast… watch:
Here’s the results:
You can even do this in bulk pretty easily, just drag the photos into photoshop, select object, invert selection, save, move to next image. It takes about 3-4 seconds per image on my now aging home computer, so even a dataset of 100 images will only take maybe 5 or 6 minutes to do the whole lot.
The result of using all those nice clean images? A wonderfully reconstructed object:
See for yourself on sketchfab:
Obviously, you should never take photos like this if you can help it. BUT, if you do want to recover a model from a less than ideal dataset, Photoshop’s new object selection tool might just be the key to getting a great model from poor photos.
Dr Peter L. Falkingham 
Interesting, thank you. A couple of comments:
1. There are a number of online “image background removers” which, from experience, work to a greater or lesser degree. I’ve used the free option on http://www.remove.bg for a couple of tasks and it’s worked quite well ( other background removers are available ). I tried it with one or two of your sample images and it seemed to work OK and worked very quickly.
2. The paid version of the website above has an API so I guess that it would be possible to automate this process. I’m not sure how well it would work for images that are not against a clean background but I’ve tried a couple of my images of standing stones and it seems to do quite a good job. I’ll have to investigate further.
An update to this. I’ve found a FOSS tool for background removal that runs locally – https://github.com/OPHoperHPO/image-background-remove-tool. From a quick test it seems to work pretty well but there is some scope for making it better.
More details here – https://theretiredengineer.wordpress.com/2020/10/11/background-removal-for-photogrammetry/.
Good to know – thanks!
I have my own method that can be used in bigger data sets than 100. Of course nothing is better than manual hand masking (or apparently that method ) but done that for 500 pictures for example? nightmare. Below is link to OLD method. New one increase exposure to really high levels (to unify background), then create mask from difference and back with exposure do normal level – thanks to that difference layer that is similar color to one in background mask without breaking original object. From my observation, those remains after auto masking are not problematic for pg process btw.
I don’t have PS so can’t make direct comparison but glad to see such tool anyway