[Contextcapture] Reality model from German Flown Aerial Photography, 1939 - 1945

To push the limits of Contextcapture there is a interest in creating Reality models from analog historical photos.

Historical background

At the end of the Second World War (1939-1945) the British and Americans discovered a mass of German aerial reconnaissance imagery, hidden by the Germans across multiple locations. The story of its discovery and exploitation is fascinating and complex.

Much material came from Hitler's mountain retreat at Berchtesgaden in Bavaria, Germany, and was condenamed DICK TRACY. An Americam team secured this only within a matter of hours of Russian forces arriving.  Upon its arrival at RAF Medmenham a British-American exploitation project began. As the project developed all the imagery became known as 'GX'.

GX is such a large quantity of imagery that preliminary sorting work went on until 1949. The intelligence discovery provided mass aerial imagery of Eastern Europe and the Soviet Union on a scale unachievable at the time and remained a key intelligence resource for more than two decades. Even by 1960 much of the British and American target data consisted of predicted radar imagery of targets which were derived from the GX images. Althought the British and Americans made a number of clandestine flights over the Soviet Union in the 1950s, the coverage obtained was small and GX was only replaced with the arrival of satellite reconnaissance imagery in the 1960s.

All of the WWII images have been declassified, but most remain in the archives in film or paper form, accessible by request but not yet online. At NARA, the collection of more than 1.2 million images is identified by archive number 306065, labeled German Flown Aerial Photography, 1939~1945, and forms part of Records Group 373 and  of the Defense Intelligence Agency. Some part of archive is accessible from UK archives NCAP and Records Group 242.9.4 as target Dossiers.


Examples:

Some of the aerial photo examples can be downloaded here  . It is from Katyn area in  year 1942/1943.

Additional metadata as flight altitude, scale, position,dat can be seen here in overlay index file 54N031E- https://catalog.archives.gov/id/44258499

Camera:Luftwaffe/Zeiss Rb 50/30 or Rb 75/30 (other cameras RB 20/30 or Rb 20/18 or Rb 50/18)
Focal: 503.16/748.88/752.7/753.6mm
Image size: 32cm(12.6in) 
Sensor size: ~30cm
altitude:8000m

gx-1562-sg-104-2012-001_a - 31.7546,54.7782 (106.27°)
gx-3707b-sd-39-2012-001_a - 31.7897,54.7790 (-80.41°)
gx-3707b-sd-40-2012-001_a - 31.7722,54.7623 (279.27°)
gx-4257-sd-120-2012-001_a - 31.8000,54.7808 (-71.14°)
gx-4257-sd-121-2012-001_a - 31.7901,54.7796 (283.71°)
gx-4257-sd-122-2012-001_a - 31.7782,54.7830 (-75.16°)
gx-4257-sd-123-2012-001_a - 31.7683,54.7820 (-75.57°)
gx-4344-sd-76-2012-001_a -  31.7965,54.7726 (280.93°)

Workflow:

  1. As start I georeferenced images in approximate locations using Google Earth(Katyn-luftwaffe-aerial-online.kml). 
  2. Then created a photo position file for import in CC from KML - Katyn-pose-ypr.csv
  3. Cropped images using same resolution so photos from one flight creates one photogroup with same resolution
  4. Create new project in CC and added photos and imported photo positions.
  5. Provided film size 30cm as sensor size and focal 

The problem is that either model is upside down after aerotriangulation(even if block type set as nadir) or aerotriangulation produces model which is on one side if using geo metadata. Only 4 of 8 photos can be used but that is expected because some photos are from different year. Adding fake control points I am able to set it in correct position but for some reason later mesh production fail.

So the question is what options in Contextcapture could improve aerotriangulation when using such data? We have succesfully produced models from other BW aerial photos but still there is room for improvement. This is great example to test capabilities and to find out best approach.

Attachment - KML file and photo positions in CSV.

Katyn-luftwaffe-aerial-online.zip

Parents Reply Children
  • We recently georeferenced batch of 200 historic aerial photos from interwar period(barn find in national archive) with total set of 15k as a proof of concept . Contextcapture is ok but artificial limitations makes it not so suitable as Pix4D. So most likely if no improvements are made then Contextcapture license will not be considered for this project.

    Contextcapture disadvantages regarding old aerial photos:

    • Does not support true greyscale images but no issues if it is saved as color image.
    • Does not allow to lower minimum number of tie points to 2 so orthophotos are with holes even where there is overlap.
    • Unreliable DSM because of bumpy roads feature.
    • No option to update original input photos with georeference(world files).

    Result:

     

    I have also found such excellent paper:

    Reconstructing the Past in 3D Using Historical Aerial Imagery (Michael Broadbent)

    Main conclusions:

    • Photos need to be standartized/normalized for use in SfM softwares. It means correction for rotation and same image size(script attached) so they are imported as one photogroup.
    • Photos need to be from same flight as changes in vegetation cause issues with AT. To be safe each flight need to be as separate photogroup. Needs improvements in feature detection(spline detection) to be more useful for such cases. 
    • Camera metadata helps. If AT is wrong(upside down/curved) it means sensor size is not correct. Need to provide only size of capture not full physical frame size. 
    • AT works even without geolocation. So GCPs can be added after AT.
    • In latest Contexcapture versions need to use "Legacy engine" preset.

     Aerial-standartize-python.zip