Photoscan with Old Aerial Photos

[sarko]

Hello,


I am using Photoscan on a couple different research projects and we seem to have hit a wall that I'm hoping someone can help with.  We are using aerial photography that is about 30 years old for which we do not know that camera calibration coefficients.  We are generating orthomosaics using 4 or 5 images at a time.  Our general workflow is as follows:

1.  Image are scanned from film at 1200 dpi (approximately 1.2m ground resolution)
2.  Crop all images to remove fiducials and black border.  Images are cropped to the same dimension so that PhotoScan interprets them to be from the same camera. 
3.  Load images into Photoscan and align images at highest resolution
4.  Generate geometry using Medium resolution height field.  200000 faces. 
5.  Apply ground control matching known points in the imagery.  Our ground control is collected using a survey-grade GPS and is generally accurate to better than 20cm. 
6.  Export orthomosaics as geotiff

I believe this workflow is appropriate.  Not knowing what the camera calibration is we allow Photoscan to calculate the K1, K2, K3 coefficients.  The values it calculates are reasonable for this type of imagery, so that seems ok. 

When we analyse the resulting mosaic (approximately 250 sq km) we find that many areas have significant deviation (> 10m horizontally).  The areas where we apply ground control are significantly better, but as we move away from those areas it seems the model is breaking down.  We have used varying numbers of ground control, but going from 3 to 4, 5, or more doesn't seem to improve the result, so we have been sticking to 3 to try to limit the warping that Photoscan is doing. 

We are hoping to generate orthomosaics that meet the U.S. NMAS 1:24000 standard, which is 12.2m at CE90.  At the current time, we cannot do this . I would appreciate any thoughts people may have that might help us improve our product. 

Thanks,
Scott

[Diego]

Hi Scott,

Actually I have 15 years of experience in photogrammetry, and I am a specialist in digital photogrammetry for elaboration of cartography, and the procedure that you raise is not the best.  I have worked with analog pictures in photoscan, but another very different perspective and have had excellent results.  The first is to perform a photogrammetric scanner at 15 micron is 1693 dpi, if you do it in a conventional scanner going to introduce more errors in procedure.  You cut the photos is a great error.  I can help you improve your workflow, if you send me the pictures and I guided in the process, no problem.

Best Regards,

Diego

[RHenriques]

As Diego Torres said, Photoscan is pretty impressive in determining DTM's from aerial photos (even without any flight or camera parameters) or scanned photos. Try to avoid any distortion produced by the scanner and use an appropriate resolution. Some old photos are not well suited for DTM extraction due to overexposing.
Cheers

[RHenriques]

As DiegoTorres said, your problem is precisely cutting borders with the same dimension in all photos. Aerial photos do not have the same border dimensions because the black border width is not the same among a photographic session strip. The only thing that maintain the proportions relatively to the photographed area are the fiducial marks. So it's not a good idea to eliminate these marks and the borders. First cut all the photos using the fiducial marks as guides for the cutting borders. After, if necessary, you can straighten the photos if you notice some need for rotation but you must ensure that the rotation is perfectly justified (use photoshop to measure rotation for instance). After this pre-processing you can then apply your workflow because the elimination of the black borders will not have the Aerial photograph border as reference but the fiducial marks, which will not add any distortion to your imagery and will allow Photoscan to calculate the correct parameters.

Cheers

[Matt]

I have had very good success with photoscan even with film scans of aerials from the 1940's. To obtain accuracies less than 1 metre I actually pre process the raw photos significantly before use in photoscan.

Within a photogrammetric package I perform an interior orientation of each photo to fiducial coordinates and remove the lens distortion (I dont do this step any longer). I then write these oriented photos back out to file and clip them to the fiducial extents so they all have the same image dimensions and Principal Point of Symmetry. I then run the alignment and geometry build then tie this surface to very accurate ground control ie Lidar/survey marks.

Highest accuracies are obtained where there is significant overlap (>40% forward and side) without significant overlap or any sidelap the mosaics tend to flare out and up at the ends, so it may be worth extending the mosaics so there is overlap too outside your area of interest.  It is also worth noting that choosing a calibration result from a photo that has overlaps on each side and applying that to all the photos works well if you have no camera calibration information.

This process has worked for full resolution scans and mosaics up to 40 photos you just need a machine with lots and lots of RAM if you want to work at the highest precision.

Hope it helps

[EvanThoms]

Hey Matt,
What program are doing the interior orientation with?
And is your cropping to fiducials an automated or manual process?

[Diego]

After conducting internal orientation is fully automatic with Intergraph ImageStation

[Matt]

Hi Evan,

I have been using a package called Landscape Mapper which actually allows you to output the oriented photo as part of a photogrammetric workflow. I havnt found too many affordable packages that allow you to do this (any thoughts diego?). With a good scan and accurate calibration report for the camera (which can be hard to track down sometimes) you can align the photo to the fiducial to within a pixel or two.  The software was written by a friend for his mothers vegetation mapping work.  I used to sell it for them but I dont think they sell it anymore. I can chase it up for you if you like.

Hope it helps.

Matt

[EvanThoms]

Thank you Matt and Diego,
I will look into those. In the meanwhile, could either of you, or anyone else, explain the problems, if any, in simply "georeferencing" the images to the pixel coordinate space by assigning the calibration-derived coordinates of the fiducials, in ArcMap for instance, to the fiducials and exporting a rectified image? Is that basically the process in interior orientation? Although, automating that process for many photos, I presume, would require some kind of image recognition algorithm to find the fiducial marks?

On the other hand, I just ran some old photos though PS after only specifying the focal length and got much better results, maybe good enough. Perhaps that's all the pre-processing I have to do; I would have to run some tests.

[RalfH]

There's a bit of a problem with the suggested workflow: If the images are simply cropped to remove the fiducial markers and image borders (point 2 in the workflow), there is a high risk that the pricipal point of the images is not exactly at the same pixel coordinates in all scans. If these images are then calibrated as being from one camera, quality of calibration (and of the 3D results) will suffer. To ensure correct calibration as a single camera for all images, cropping has to be done in a way that all fiducial markers are at the same pixel coordinates. Only then "group" camera calibration is appropriate. Fiducial markers should therefore remain in the images and can be masked in Photoscan to avoid interference with 3D processing.

[spectrageo99]

Hi Evan,

I have been using a package called Landscape Mapper which actually allows you to output the oriented photo as part of a photogrammetric workflow. I havnt found too many affordable packages that allow you to do this (any thoughts diego?). With a good scan and accurate calibration report for the camera (which can be hard to track down sometimes) you can align the photo to the fiducial to within a pixel or two.  The software was written by a friend for his mothers vegetation mapping work.  I used to sell it for them but I dont think they sell it anymore. I can chase it up for you if you like.

Hope it helps.

Matt

Hello Matt,
I've unsuccessfully searched for this "Landscape Mapper"  application, seeking to use for performing interior orientation (defining historic camera calibration geometry, inputting fiducial coordinates, and measuring fiducial coordinates).  Any suggestions in where this may be found?  Any low-cost alternatives?
Thank you,
Dane