Hi. I would like to ask some things regarding fspy and blender, before starting I would like to do it in the right category within the forum in order to get the best of it and not disrupt the forum.
Thank you
I switched the category to Help-Other software with a tag of FSpy…that should do it
So please go ahead and ask your questions…
1 Like
Well if He’s not going to do it, I will. lol
Fspy is pretty great, but some images are just plain impossible to match. Doesn’t keep people from asking for me to render their product in that image though.
You can’t really account for lens distortion can you? If so, do you have to “undistort” the image first, then bring that into fspy?
The one that’s Really difficult is if someone shifted the camera back, or Worse, make it Look like they did that Afterward in photoshop. That’s when it becomes impossible.
I did a quick search, but didn’t see if there were somewhere to have feature requests.
I can’t remember the software I used. It Might have been an Autodesk acquisition at some point, so I can’t recall. But they had a decent camera match tool, but it let you pick several perspective lines, then gave you an average for the convergence point.
I would love to be able to add a few more Y lines, and a few more X lines there. Even more awesome if it could show you the worst line, to know if maybe you just picked a bad object, or if you need to tweak your points a bit. I mean, you move just half a hair with one single point and you can be pretty far off. It all depends on the lens used of course.
But multiple lines would help So much, especially if the scene is a bit more organic. You could be kind of Guessing at the lines and it might still give you halfway decent results, because guessing over and over all around the image, you’re likely to get somewhat close to converging lines.
Of course, I’d also like to enter the lens length if I already know it, which might help.
I match the camera. Then after importing I model just a little of the scene quickly, until I get to an object that I am pretty confident I know the size. Parent everything, including the camera to that object, then scale it up to the correct size. Then unparent and freeze transforms.
Yes, fspy has a scale tool, and it Kind of works. But it only works on axis, and you’ll find it’s much easier to have the axis on a ground of some sort, whether that’s a tabletop, actual ground, roof, etc. I usually place it at the base corner of something. For example, I just matched this one:
I noticed there are cookies and…jam? So definitely not an image from the US. And the table’s legs are jacked. (pretty sure those are ikea legs on that table, which seems warped, but could also be distortion from the lens).
Most rugs in the US are in increments of 6 inches, and Mostly 12inches. It didn’t seem to be lining up to any reasonable size. I actually matched to a copy of the image I added a Bunch of contrast to. If you’ll notice, there are virtually no visible lines in the image that go in X direction, but if you increase the contrast to hell then you can make out a few of the tiles.
And yes, if it were in the US, the tiles would be a standard size…but again, not the US.
Thank goodness for the lady in the wheelchair. I made a 24" diameter cylinder and voila, scale “matched.” That put the rug at 53.2x72.9-ish.
You may also notice I have two cameras. The one selected is a standard match, and the other one is shifted. There isn’t much difference between the two, so if that’s the case, use the non-shifted one.
I’ll probably end up using this one in keyshot, because we already have the product there ready to go. And the cameras there don’t match (SO Dumb). So you have to invert one of the shift numbers. And there have been render issues with shifted cameras as well.
fSpy it is a very simple tool for compositing CG with right perspective, if you need to do precise work, there’s other softwares like 3DEqualizer, PFTrack and SynthEyes, which are tracking softwares. For a proper workflow you will need lens grid and camera sensor size. Check VFX Tutors channel for 3DE tutorials teaching matchmove, survey and lens distortion solve for production.
Do any of those actually match a single image? Because tracking software usually doesn’t, and needs multiple images in order to create a camera. You can do that in blender also. But since I’m using stock photography, it doesn’t much matter.
And sure, using a lens grid to calculate distortion would be amazing…if I were in control of who is taking the camera.
But again, not doing animation, and i only have access to a single image.
I think what I used before was Autodesk Imagemodeler (discontinued).
And fspy Does work fine for what I need. I would just love to have a few more features. They aren’t particularly Advanced features either. I mean, Adobe is using AI (poorly) to determine a camera, lighting, sun location, etc for their software. I’m just wanting to add a few more lines and use a farily basic math function to average their solution. I mean, I suppose I could just do that by drawing a bunch of lines and exporting a bunch of cameras, then averaging them once in blender, but that seems like all too much effort for something that would only be Slightly better.
But that’s the kind of solution that works better for images you don’t have control over.
Sure, if I’m directing the photo shoot, I have that camera on a tripod, take some images of specific objects that are super easy to match first. I make note of the lens and camera, and I make sure that if our photographer wants to use a zoom lens, try and stick to even increments to make it easier for me to match if it’s not a lens that records the zoom length in the exif data. I also would avoid using any lenses wider than 35mm whenever possible to minimize distortion.
I do, guessing, 40 of these images per year? And only a few of them are for clients. So Definitely not something we would be paying several hundred per month for software licenses for.
Just thought I’d add some tips for fspy, and see if anyone knew of some of the answers to questions I had is all.
Shifted camera / cropped images can be painful indeed, but it’s still possible to determine the principal point position.
I created a tutorial on how to do this (and not only this, as it covers all fSpy options) for 1-point, 2-point and 3-point perspective (as the strategies differ).
It’s available here https://www.youtube.com/watch?v=daiMOYR8GS8
I hope it helps
best,
bartek
1 Like