Forum : Photosynth Lounge

Photogrammetric, Oblique Image Stitching, Pets Dressed in Clothes Photos… this is a place to chat and share stories with your fellow Photosynthers. Not all topics have to be about photography, this is a place to relax and chat about whatever you fancy.


Topic: What does Photosynth want?

Report Abuse Report Abuse
tbenedict (Over 1 year ago)
The topic for this thread probably looks a little whiny, so I'll clarify:

I just had my first completely failed synth.  10GB of images, and it finally ran out of memory.  (Hey, I had to TRY!)  But it occurs to me that Photosynth may not be using the full 4000x3000 size of the images I've been tossing at it.  I could save a lot of processing and uploading time if I re-sized to the max size Photosynth could make use of.

But... er... what IS the maximum image size Photosynth will actually make use of?

Along those lines, I've been working on a semi-automated pipeline for aerial images.  Tossing a re-size in there is no big deal.  But this raises some related questions and points:

(more)
tbenedict (Over 1 year ago)
1 - False color images made using a principal component algorithm doesn't cause Photosynth to barf.  Since the axis with the most contrast winds up in luminance, typically, that makes sense.  (This is what got me going with the pipeline in the first place: doing automated false color rendering to look for variations in rock and vegetation.)

2 - What would image sharpening do to Photosynth?  Any time a re-size is done, images tend to lose accutance.  So a slight unsharp mask can make them LOOK better, even if it doesn't actually add information to the image.  This is also pretty straightforward to do, but I don't want to mess up the features Photosynth is trying to match on.  Any idea what sharpening the source images does to a synth?

3 - I'm tempted to do some sort of auto-contrast or auto-levels as well, but this does directly mess with luminance.  Any idea how Photosynth deals with this?

(more)
tbenedict (Over 1 year ago)
4 - I'm planning to use PTLens to remove barrel distortion and chromatic aberration from my images (the CA on this camera really isn't bad, but why not?)  PTLens is pretty consistent, but is this just creating extra work that Photosynth is going to take care of anyway?

The one piece that's missing from all this is the blurry frame removal.  It's a fact of life with kite aerial photography that at some points the camera may focus on the kite line rather than the scene.  I don't lock my camera at infinity focus because sometimes I really AM flying low, and want the close up bits to be in focus (e.g. my lighthouse synth).  But I haven't found a program that will easily do this.  Still working on it.

Any other thoughts I should be thinking of?

Tom
Nathanael (Over 1 year ago)
Tom, my reaction to the first part of your post is that your out of memory error was probably not due to the image tile generation (the part that deals with the full resolution images) but rather the image feature matching. (Is this right? Could you check your log file and find out?) 

While it is true that you would save time uploading by using smaller versions of your images, all image feature detection is done using 1.5 - 2 megapixel (the reports vary) versions of your images, so the processing load should be identical for the image feature detection, matching, and scene reconstruction with any photos 2 megapixel and above, if I understand it correctly. 

Here are a couple of links along that line that I based what I said above on.
( http://getsatisfaction.com/livelabs/topics/what_image_quality_and_resolution_should_i_use ) 
( http://getsatisfaction.com/livelabs/topics/whats_the_best_size_resolution_for_taking_pictures )
tbenedict (Over 1 year ago)
Nathaniel, I think you're right on the money.

I tried a new technique on a recent set of synth images, and while I think it helped things, it looks like it helped a little too much.  On an earlier photo session at Kiholo Bay, I ran into problems because I didn't have enough variation in camera location when photographing the scene.  So I wound up with three planes of features in the point cloud.  Very weird.  I took everyone's advice and did the next session from a number of vantage points.  This did a good job of fixing things.  The point cloud was very clean.

If a little is good, more is better, right?  Here's what I did:  I pointed my camera 90 degrees to the kite line, tilted it down at about 45 degrees, and with the camera set for continuous shutter mode, I held down the shutter button while letting out kite line at a nice steady pace.  That made for long runs of images with the camera basically moving sideways.

(more)
tbenedict (Over 1 year ago)
I was doing this in an effort to make stereo pairs for doing aerial stereo photography.  But I think it also made for truckloads of matching points between images.  The point cloud from that synth was very very dense:

http://photosynth.net/view.aspx?cid=bcab64f4-054d-404c-98e9-eb6188a0cdde

I did that technique on another set later in that day, and kept having failures with that one, too.  I finally reduced the image set to just a few hundred, and it worked.  But I've done 1000+ image synths with no problems before.

I finally watched it, and yeah, it was during image feature matching that it was dying, just like you thought, Nathaniel.  So reducing image size probably wouldn't save me much grief.  Ah well...

(I think I just need to get a little less shutter-happy with my camera!!)

Tom
Nathanael (Over 1 year ago)
:) One more variable to the equation is that the number of images is actually independent of how many synther-friendly image features there are per image. 

1000 photos or more of one subject could potentially have many fewer scale/lighting/angle invariant features that the algorithm is set to recognise than several hundred of another subject more synthy like rock (or lava?). Granted you may have done other 1000+ image synths over rock/lava before (I'm not seeing them in your public synths at the moment, though) but recognisable image features are bound to vary with location/focus/exposure.

Don't let my hypothesising put you off shooting lots of coverage. I hope you get answers to the other questions you had.

Nathanael
madeeds (Over 1 year ago)
Since you have the pipeline pretty much set up, it would be awesome if you could try some controlled experiments with a couple data sets to see if the image processing hurts or helps Photosynth.  I'd love to see that.  Just take a couple sets of source images, keep it down to about 300 images each.  Then try adding or eliminating stages to your existing pipeline and see if it makes any difference.  To check your results, run Fiddler while viewing the synth and see how main .bin (point cloud) files it downloads.  For the synth you link above I see 134 point cloud files, and each is about 96Kb.

BTW, I love your KAP synths.  They are an amazing vantage point.
You need to be Signed In to add a comment. (Are you new? Sign Up for a free account.)