Get all your questions answered about our latest Photosynth Technical Preview.
Looking through some of the Preview content, I've seen some fairly painful synth playback/reconstructions. This may be the fault of the photographer, however I'm mainly looking for cases where Photosynth's viewer seems to be at fault during playback.
If you find any where Photosynth is at fault, leave a link here and a note of what you see wrong with it.
Some really big gaps between photos here during the slideshow.
Noticeable gaps between photos again.
Gaps between photos again;
Bizarre low resolution red and yellow pixels against black a little way through the slideshow
Does okay at the beginning, but ends in madness.
Beautiful photography for the most part - much better than mine. The reconstruction leaves something to be desired.
You can see the outlying camera positions in the map view and they make for a pretty glitch video preview.
1. Looks like it has been deleted.
2.Not a single smooth path. The photographer just "sprayed" the tree with photos, and the algorithm tried to move in shooting order, leading to big jumps. Until we have 2-dimensional navigation (as opposed to the 1D nav we currently have in the new technology)we're going to get bad experiences out of this kind of shooting.
My attemps so far have been large blocky gray areas. When doing panos I zig-zag up and down while rotating.
Here's an example: http://photosynth.net/view/84ac47fb-9e50-4ba0-8bb8-89b6e805e17e
Here's the same set of photos in a Pano
and in a old Synth (plus some more cloud shots)
@dhaines, there was a related discussion on panoramas shot for ICE here if you're interested:
Apologies if you've already read it.
Nine months ago, David Gedye on the Photosynth team suggested:
"What do you guys think about a hybrid approach: Low res stitch on the phone, but upload full frames and get a hi res stitch done in the cloud a minute or so after upload? Note that hi-res uploads would be too big do over 3G, and probably bigger than you'd want to do over 4G, so the scenario is more like: upload later when you are on Wifi. Thoughts?" Source: http://bit.ly/psfhrp
Also related: Joscelin's comparison of single strip panoramas in Photosynth 1, ICE, and Photosynth 2:
Note David's reply there too.
I got interrupted a few days ago when replying to the initial set of 5 samples. Here's the answers for the rest.
#3: This is the common scenario of a "2D wall", which we do not yet support. The photographer has gone back and forth, covering the group in the hope of getting a full reconstruction. We actually DID do a full reconstruction in the backend, but as you know our viewing path is only a single curve, and tries to follow the shooting order. The big gaps come from moving from photo N to N+1 across a reconstructed area. Because our viewer only uses photos N and photos N+1 to render this transition, it couldn't really show anything meaningful for this big jump. Bottom line is that for this kind of input, the old Photosynth is much better.
#4. Should have been shot more densely, but even so it looks like we have a reconstruction problem right at the end. (See "m") Terminal photos that are significant jumps from others aren't recommended.
#5. This one is interesting. The glitch you see here is caused by a very common shooting behavior that we generally can correct for, but couldn't in this case.
To see this behavior, hit the "m" key and look at the top of the arc. Note how there is a dense "outer" set of photos, and a sparser (but quite sufficient) inner set in that little region. This is case of what we call "overcompleting" the circle, and it is so common that we have logic to detect it, and ignore frames that seem to be part of such a "tail".
That logic only works when the sorting of the "DateTaken" field in the EXIF agrees with an angular sweep sort from the center of the reconstruction, which it will if people take the photos in normal spin order. Here's what Mike says about this one:
It wasn’t triggered because the time ordering was deemed unusable. There are multiple images with timestamp 1387624011 (i.e., 2013:12:21 11:06:51), and the timestamps are not monotonically increasing.
This one has a great deal of funkyness in it. I have a feeling its because my speed and altitude changed frequently also the texture of the rocks and grass was a little too much. Watch the sign leap up at the end ;-)
Several weird things going on with this one: http://photosynth.net/preview/view/a05a7f03-d315-4ac9-8427-21811cd597cb
- you have to move up and downwards even though it's a panorama
- the interpolated path looks out of place
Pineapple spin synth with too much texture in the stationary background but the play speed is out of control fast.
Walk synth begins great, but photo projection falls to pieces at the end
Photo projection falls apart going in to the turn at the far end of the room
It's like the virtual camera just falls off the path here, leaving the geometry to explode into the viewer.
13 - 19: http://bit.ly/psfboop
Don't Panic asks about several of his synths' reconstructions here:
Just curious why the first half of this spin didn't synth in.
Someone turned out several lights for photos 5 and 6 of 13 in the middle of this spin, breaking it at that point.
I've tried adjusting the histogram of the two darkened shots to help them match in and resynth, but it was no more successful.
Critical fail - check map. This is a almost strait line.
This ought to have been a simple 160 degrees or so spin around my brother playing Total Annihilation against our cousin but the left hand side is folded into the center.
re 22:Looking at your input files, there is something really weird with the DateTaken field of your EXIF. It doesn't proceed in time as we'd expect -- it goes back and forward between 3:56pm and 3:57pm for most of the sequence. That's what's making our synther confused because this is the signal we use for order. Any ideas why this might be?
24 + 25: http://photosynth.net/discussion.aspx?cat=00581351-82d8-438d-a37b-7eadb3fb4991&dis=7cb08b9a-6295-45d5-92d4-15cc87ef2f78
Andy C 1202 asks about the reconstructions of two spin synths which fell apart.
My spin synth of a windmill at a local tulip festival smashed to pieces here: http://photosynth.net/preview/view/5b7acef5-5e2a-4f8c-843a-b1f46acae754
I know the angle that the ground was seen was very shallow and often occluded by passersby, but I had nearly 50 input shots which I hoped would be enough coverage to compensate.
I fear I may have spiraled outward when coming back to the beginning point as well, but the reconstruction is far too mangled nevertheless.
Take a look at the map Nate ('M') there is a massive gap at one end (where bits fly everywhere!) and some sort or loop at the other. Has PS dropped a lot of shots at the start - and have shots been duplicated at the end?
Yep, I checked the map.
I took and uploaded 49 photos and there are that many in the reconstruction, mangled though it be.
There may be some overlap in angles when I completed the circle but as I said, I think I was further from the windmill at the end than the beginning so it isn't sheer duplication.
I'm currently trying the set in Photosynth 1 to see whether the results will be different.
Here's the synth in Photosynth 1: http://photosynth.net/d3d/photosynth.aspx?cid=35a0bfd0-fa56-4670-9b92-431b94de2ec0
There's still a noticable error in the reconstruction, but it's at least headed in the right direction.
Here's a screenshot of the PS1 camera positions from overhead: http://1drv.ms/1poWCOg
Here's one that just went totally crazy.
This one should have been an outward spiral which straightened out at the end with a view of the city.
The navigation is backwards in this synth: http://photosynth.net/preview/view/24f93c5e-902f-4fa3-8318-ecec29b8181f