[OpenDroneMap-users] processing undistorted Go Pro imagery

[Aaron Dennis]

Hello,

I've seen some incredible output from example datasets with OpenDroneMap.

Right now, I'm working with some researchers who fly a UAS with a GoPro
Hero 3+ Silver camera to collect video. I'm trying to pull image frames
from that video, remove the fisheye distortion, and run the images through
OpenDroneMap.

I pull frames from the video with this command:

$ ffmpeg -i video.mp4 -r 1 %03d.jpg

Then, what I've been doing is removing barrel distortion (fish eye) from
the images with $ imagemagick:

convert image.jpg -distort barrel '0.030530 -0.124312 -0.038543'
corrected_image.jpg

I also add exif tags to the imagery using exiftool:

$ exiftool -make=Gopro corrected_image.jpg
$ exiftool -model=HD3 corrected_image.jpg
$ exiftool -FocalLength="2.8 mm" corrected_image.jpg

My challenge now is determining if 2.88 mm is still the correct exif tag
for the image, having already successfully removed distortion. Also, I'm
not sure how this will affect the sensor size in ccd_defs.json. I've tried
running the images through OpenDroneMap both distorted and undistorted
using the exif tags shown above, and each time I end up with an empty point
cloud.

Matching key points appeared to go okay.

   14151 matches (23.8%) took  6.63s for
/home/ubuntu/images/reconstruction-with-image-size-2400/matches/0-1.txt
...........

   14093 matches (23.7%) took  6.93s for
/home/ubuntu/images/reconstruction-with-image-size-2400/matches/0-2.txt .

    1670 matches ( 2.8%) took  6.94s for
/home/ubuntu/images/reconstruction-with-image-size-2400/matches/0-3.txt
...............................

   17644 matches (28.4%) took  7.22s for
/home/ubuntu/images/reconstruction-with-image-size-2400/matches/1-2.txt .

    1762 matches ( 2.8%) took  7.20s for
/home/ubuntu/images/reconstruction-with-image-size-2400/matches/1-3.txt
.........................................

    1901 matches ( 3.0%) took  7.18s for
/home/ubuntu/images/reconstruction-with-image-size-2400/matches/2-3.txt

But running pmvs seemed to maybe filter out all those key points? Running
meshing returned an error that the point cloud was empty.

I was wondering if anyone has a solution to processing Go Pro imagery or
video or if anyone has some ideas about what focal length or CCD size would
be appropriate for undistorted imagery.

Thanks,
Aaron


*Aaron Dennis*
aaron.p.dennis at gmail.com
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.osgeo.org/pipermail/opendronemap-users/attachments/20150416/35d9cbbb/attachment.html>