Driving a car and getting sequential shots

[skinkie]

After seeing several issues on meshroom I have been collecting a driving dataset myself.
alicevision/Meshroom#846
alicevision/Meshroom#895

I first validated that results by capturing 1080p@15fps is similar to high resolution photo's. Obviously less detail, but it still reconstructs. I then did a 1 out of 4 frames export in ffmpeg.

ffmpeg -ss 2 -t 60 -i ../VID_20200607_163932.mp4 -q:v 1  -vf select='not(mod(n-1\,4))' -vsync 0 images-%05d.jpg

Doing so I get rather ok results (111/161 photo's) but it is not always clear for me why a previous or next photo doesn't match. I would still be very convient to learn why two consecutive photo's won't work.

I have achieved these results using:

• VoCTree minimal 0
• NbMatches: 100
• Sequential NbMatches: 20

SfM:

• Estimator: Loransac (Don't know why...)
• Observation Constraint: Scale (Don't know why)
• Max Ransac: 40 seems to be optimum
• Min Angle For Triangulation: 1.5 (descreasing it further gives less reconstruction)
• Min Angle For Landmark: 1