Do you have an idea for an awesome feature we should add… or hate the way we’re currently doing something? Share your ideas and suggestions here.
I get the feeling that the power of the synth client has been capped to ensure that the detect time is minutes not hours for a small set of photos.
I would like a way to set how much detail (and how long) the synth client goes into when scanning the images pre match. The more detail the better the synth, right?
I have faced issues where even though images are of the same building or object, they do not get detected - I believe that if they had more time to scan the images and were able to produce more points/patterns then the synth would be better.
I am asking for an advanced more which allows a user to select the intensity of the synth.
Related to this idea is this one that I shared a while ago:
A sketch of a hack to get around the fact that the synther uses lower resolution versions of the uploaded images for the image matching and positioning can be found here:
To be perfectly honest, though, it's quite simple to practically guarantee that the synther will put groups of shots together by providing a semicircle of shots around a given surface (teaching Photosynth to recognize it from any angle that it can be approached from) and taking shots looking and walking from one surface to the next.
Although the hack that I link to above should allow you to extract the maximally dense pointcloud that the synther is capable of generating from our photos today, using original images plus crops of them is a messy and what's more, time consuming thing, not to mention that it's likely that that many extra features is far more likely to crash the synther unless you're beginning with very few photos indeed or using fairly large crops of your photos which will also be downsized by the synther.
I'd had the theory since at least September 2008 but it's taken me until this past Friday evening to actually make myself begin to test it. My original synth was 130 images and 100% synthy. The hack version uses 1000x1000 pixel chunks of the original images (12 crops per original) as well as the original photos themselves, just to tighten the alignment up on all the crops. This yielded a total of 1,690 separate JPGs and the synther has been running for about 35 hours now.
Ah, bummer. I'll have to try my cheat with fewer input photos. My test synth died this morning after about 50 hours and 14 minutes of pure computation and that computer has to be run in OSX during the day for work.
I might have to wait until Friday night again for another big test if that first timeframe is anything to judge by...
Or maybe I'll just try it without so many photos this time. I kind of hate to do that, though.
As usual Nathanael, you've taken the bull by the horns; here, by both informing to such an extreme level (a typical move by you) and also in your testing and explaining of methodology.
I feel your pain about that 1,600+ photo synth using the "Crop Approach" to improve and perfect synthiness. My sense is such an experiment should almost assuredly work wonders, once you find the correct number of photos that will allow it to process without failure. The low limit on the number of photos that actually work in the Photosynth synther is the second most frustrating aspect of Photosynth for me. (The fact that the problem occurs after hours and hours of processing most of the time is first, and the incorrect aspect ratio of the thumbnails as third, probably)
Anyways, impressive work, Nathanael!
Well thanks, inveritest,
As far as the number of photos that the synther can process, I've had it succeed with larger numbers of original images than that experiment above used. I still haven't seen a single synth which successfully organised 2,000 or more full quality photos (2,000 or more low quality video frames: yes), but 1,800 or so seems like a fairly common limit, provided that you've spread your coverage out fairly evenly and not focused too long in one spot when covering large areas.
You may well have already seen this topic, but for others' sake, here is more on what sorts of things are limits for the synther (which leads one to imagine how to maximise its performance):
In any case, my last reply is taking the thread off topic, so I'll stop.
I don't think I'll have time this weekend to repeat my big test again, but will post my results back here if|when I get it to succeed.
For anyone interested (mwiththeat, some of your synths have few enough original photos that they shouldn't face the long computation that mine did).
A simple way to batch generate square tiles of any given size is to use the command line deep zoom tools. You'll only want the tiles from the highest numbered folder inside each converted image's tile pyramid as this one is the original resolution divided into the tile size that you have chosen. Feel free to delete the second highest numbered folder + below in each tile pyramid.
Get the Deep Zoom Tools here:
Read the readme + use this string: Dzconvert.exe source\*.jpg dzimages /tq:1.0 /ts:1000
It specifies JPG tiles, full quality, 1,000 x 1,000.
There's also a variable to control the number of pixels that each tile overlaps its neighbouring tiles which you may want to investigate, but I left it at the default of 1 pixel overlap because it wasn't going to change the number of tiles for each image.
If you want a different tilesize than 1,000 square, substitute the "1000" in /ts:1000 for whatever number you prefer.
If anyone out there feels like coding up a tool that does this reliably without generating the extraneous smaller resolutions' tiles or knows of a tool that does this well, please feel free to let me know what it is here.
A caution for those with higher resolution source images: you may need an extra layer of bigger tiles between your smallest and your original image in order for the synther to match them.
I have seen 10 megapixel images matched without need of intermediates, but your milage may vary.
After another unsuccessful attempt like the one above, I did finally make a synth with square tiles that were half as wide as the shortest dimension of my photos.
If anyone is interested, here's the link: