Loading Profile...
Torture testing and performance
Hey guys!
I didn't heed the warning and made a synth with 328 images. The synth was created fine, but I got the error A000000B people seem to be seeing. I was trying to do a "mini" torture test of the synther, but I guess I'll have to try again to see the results.
What I mean by torture test is:
-grainy and crisp photos
-blurry and sharp photos
-photos with high and/or low contrast
-dark and/or light photos
-photos with and without flash glare
-photos with little and much overlap
-photos with only hand-painted text in common
-photos with only a vibrant colour edge in common
-photos with little 3D geometry to them
-photos at sharp angles (i.e. looking down the hallway)
-photos with common features that are not the same thing
It's of the Carleton University tunnel paintings, well, only about 10% of them because there are a lot and it takes so many photos to connect them, but if the scale can improve to 3000 photos, that should do it. It's mostly photos of one side of a hallway with very colourful paintings covering it. I'm guessing that it won't synth too well (since it's partly designed not to), but it would be very cool to get a good synth of the tunnel paintings. I can think of how to adjust for some of these, especially in the context of the tunnel paintings. It could actually be a relatively easy synth if there was special handling for mural-like synths. Most of the photos are very different from each other (since it's along a hallway), and the groups that do match have vivid colours to match to. I'll report back when I've got the results.
I've also got a very counter-intuitive idea to cut down the upload time (enough to make synthing the bottleneck with 6.1MP photos on a wired connection) and so it should reduce load on the server and client side. It would require some minor changes on both sides, though, so I'll discuss it over email. I made some performance observations on the other parts too (e.g. only one CPU core gets used for reconstruction and generating the synth files, there were 40GB of page faults in the image matching, and lots of CPU use sawtooth in generating the image tiles) but it looks like uploading is a big bottleneck right now for 6.1MP images. I should really down-sample to 1.5MP as David mentioned in another thread, since most of the photos aren't sharp enough to merit 6.1MP anyway, and that would also make synthing the bottleneck.
Anyone else have torture-testing ideas? :D
(P.S. I'm not a tester, just a developer with crazy ideas who's very bad at photography, hehe.)
I didn't heed the warning and made a synth with 328 images. The synth was created fine, but I got the error A000000B people seem to be seeing. I was trying to do a "mini" torture test of the synther, but I guess I'll have to try again to see the results.
What I mean by torture test is:
-grainy and crisp photos
-blurry and sharp photos
-photos with high and/or low contrast
-dark and/or light photos
-photos with and without flash glare
-photos with little and much overlap
-photos with only hand-painted text in common
-photos with only a vibrant colour edge in common
-photos with little 3D geometry to them
-photos at sharp angles (i.e. looking down the hallway)
-photos with common features that are not the same thing
It's of the Carleton University tunnel paintings, well, only about 10% of them because there are a lot and it takes so many photos to connect them, but if the scale can improve to 3000 photos, that should do it. It's mostly photos of one side of a hallway with very colourful paintings covering it. I'm guessing that it won't synth too well (since it's partly designed not to), but it would be very cool to get a good synth of the tunnel paintings. I can think of how to adjust for some of these, especially in the context of the tunnel paintings. It could actually be a relatively easy synth if there was special handling for mural-like synths. Most of the photos are very different from each other (since it's along a hallway), and the groups that do match have vivid colours to match to. I'll report back when I've got the results.
I've also got a very counter-intuitive idea to cut down the upload time (enough to make synthing the bottleneck with 6.1MP photos on a wired connection) and so it should reduce load on the server and client side. It would require some minor changes on both sides, though, so I'll discuss it over email. I made some performance observations on the other parts too (e.g. only one CPU core gets used for reconstruction and generating the synth files, there were 40GB of page faults in the image matching, and lots of CPU use sawtooth in generating the image tiles) but it looks like uploading is a big bottleneck right now for 6.1MP images. I should really down-sample to 1.5MP as David mentioned in another thread, since most of the photos aren't sharp enough to merit 6.1MP anyway, and that would also make synthing the bottleneck.
Anyone else have torture-testing ideas? :D
(P.S. I'm not a tester, just a developer with crazy ideas who's very bad at photography, hehe.)
I’m having fun torturing the synther
2
people like this idea
I like this idea!
Tell me when this idea gets some attention.
The more people who like this idea, the more it gets noticed.
The more people who like this idea, the more it gets noticed.
-
Inappropriate?Niel, I'm glad you're "kicking the tires"! I'm shooting everything at 14 MP and not downsizing. It's mainly because if they get this thing working better, I don't want to have this nagging feeling that I need to re-upload the images in their original size. I still wonder about resolution and zooming in all the way. Has anyone one done testing to see any difference between the original high res (8MP or above) image and something sized down to 1600x1200 (or anything less than the original). If it was Photoshop, I could put the images side by side and the difference is easy to spot. I could test this, but I’m getting backed up on all the REAL stuff that I want to process and it’s not all working
The upload bottleneck is not a problem for me. Uploading is a known factor and if I have to, I can create smaller test synths just to get the stuff on the MS servers, upload all night long etc, but I know I can get the images there. With an error that completely stops me like the error A00000B that I’ve got several times already, that’s not a bottleneck, that getting STUCK! The process of renaming resynthing and waiting hours just to see another error or stuck publishing (that never shows up) is getting frustrating.
This is still the coolest thing I’ve seen come along in a very long time! Keep up the good work Microsoft!
I do feel more like a robot than a photographer when taking these photos. Maybe I can rig up a geeky looking beanie with a camera on top and I will be a walking “Google street view collector”, with the difference being that I can go anywhere 2 legs will take me. -
Inappropriate?Well, I got it to work on the third try, and here's the result: http://photosynth.net/view.aspx?cid=a...
It did slightly better than expected, but the major breaks down the hall show that it really doesn't seem to match too well based on colour, or with angled shots, or with photos having similar-but-different features (like the two different tunnel map signs, one having different arrows and building names than the other), or without lots of overlap.
I actually can't tell why it didn't find that one of the one-off photos (the one of the guy pushing on the "pull" door) has lots of overlap with two other photos, which would have connected two chunks. I also have no idea how it somehow did a great job matching the photos of the extremely blurry "magic eye" painting (in fact, I don't really know how that painting got made like that) but didn't match all of the photos of the bright yellow painting of Lilo & Stitch right next to it, leading to another break in the hall.
It never managed to connect the two sides of the hall, (except where it incorrectly matched the two different map signs), but that's pretty tough; I didn't really expect it to figure that out.
Regardless, pretty cool stuff.
I’m excited
