since I am doing photomapping of terrain,
I have experienced, that the light tends to change progressively as I shot photos.
The problem is, the flight is following zig-zag or spiral shape and as a result I have gradual variation
of brightness and contrast across ortophotomap.
In my case I am striving for fast response, not for HDR quality, and it might be difficult to repeat a mission
that takes up to 2h of UAV flight.
The question is, are there any tools that can linearly/gradually or maybe automatically apply some correction to a list of images?
We are talking abotu 300-3000 photos per flight, up to a few times per day. A real mill.
I am aware of ImageMagick toolset, but it is not a solution:
The problem is that any automatic method I know gathers statistics over a single image, then applies the correction to one or many.
Also ImageMagick scripts I tried to apply by merging photos etc show insufficient processing speed of Image Magick (it is general tool, with dozens of intermediary memory copies/logic layers during processing a script).
What is needed is a tool that analyses 3..tens..all neighboring images (consecutive in the list) and applies slowly changing correction.
I know that somebody will advise fixing exposition in the camera but thsi is NOT the way to go in 50% cloud cover changing the lighting 8 times during the flight.
What is needed is something like Deflicker algorithm used in movies, but this would be ideal but I doubt it exist for photos.
At the moment I am looking for anything that works.
I know the issue could be known to ppl doing Timelapses but I have no idea what are the solutions (I guess they usually merge the results into a movie, then apply Deflicker filter in movie processing soft, but the resolution, compression and quality is lost).
At the moment typical application of autobalance or histogram equalize/anything leads to photos where 1 out of 20 pops up with very off colors or other parameters.
Comments
I have tried Grass for 1 week 4 months and found myslef in deep Forest. The problem is that I can do interfacing on my side, preprocessing, but the core data processing fucntionality I want to leave for the others. What is interesting to me is to removing minor obstacles in almost-working solution either by UAV mission planning, autopilot export processing etc.
@ Brakar : When I was working on that project, I tried Grass (on many differents modules). My conclusion was it's very powerfull on many functions, but a lot of remaining bugs (or difficult to understand parameters ??). Really time consuming and not enough to a complete toolchain.
Documentation was poor and GUI very complex 2 years ago, but Grass is still in development, perhaps it's easier today !
GRASS GIS can probably do the blending (also); http://grass.osgeo.org/gdp/html_grass63/r.blend.html
The problem that remains to solve is to line up the individual photos correctly, and to export the image coordinats to a propor file-format. For this, I think the use of algorithms like SIFT or SURF will be neccessary.
I have a collegue who works with thousands of photo's and stitches them together:
to enhance pictures to give them the same "look": lightroom (aprox $300)
to mosaic photo's photoshop (aprox $700).
Personally i find photoshop not a product for an amateur but hey who am i (townsfool?)
These adobe products dont have a problem with photos with a size of 18 Mb and can calculate in batch.
In lightroom you work on 1 photo, select the photo's that will have the same treatment and press go.
Yes your true about tie points, they become GCP at the really end of the treatement.
Of course you can export tie points in 3D, in a first step.
There is another step in wich very dense point cloud is generated after 1st pass orthorectification is complete, but very very long step. Everything is possible I think with a so huge amount of data, but for the second pass I believe you'll wait for weeks if you need the top quality 3D reconstruction (final step) !
What are your overlapping proportions ?
@Rocca: Well I saw some results where GPU SIFT was used on large number of images that hasn't any kind of georeference information. Its speed is impressive. Almost 1 image (5Mpx) per second. Anyway, there are two working options for me:
1. 640x480 px and several thousand, sometimes even more than 10k images (thermal camera)
2. 12-16Mpx and approx. thousand images
Can you export also a point cloud of so called "GCP". I would like to name those points as tie points because they are not GCPs due to lack of measured real world coordinates as far as I understand your process of course.
@Lojze : There is no GPU support at the moment, and honestly I don't know how to do it ! Do you plan to do real time processing ? Because the step to detect GCPs between all the pictures is not the longer one (25% computing time approx, if I remeber well) with the CPU.
About the total size of final blend, there is just no limit (there's a trick of course...) ! What final resolution and initial number of picture in your work ?