auto-detect flight overlap from uploaded raw imagery

[cmosig]

From Chat: There will be more bad data like this uploaded. Given that it's visually not trivial to tell whether the overlap was good, we might need some automated warning. ODM exports a ton of metadata. I wonder if we can use some of that to create a quality flag.

report.pdf

[linear]

DT-109 auto-detect flight overlap from uploaded raw imagery

[cmosig]

I think the way to go is to read /opensfm/stats/overlap.tif, which reports how many raw images each pixel has seen. I'd read that and then compute mean and standard deviation. Note that nodata is 255, so you have to mask that out.

Here is a code snipped for this:

import numpy as np
import rasterio

x = rasterio.open("opensfm/stats/overlap.tif").read()
x = x.astype(np.float32)
x[x == 255] = np.nan

mean = np.nanmean(x)
std = np.nanstd(x)

A bad/low mean value is around 3 (obtained with 70/70 overlap) and a good value for the mean is around 10 (obtained with 75/90 overlap), I'd say based on my own tests.

We could think about generally rejecting orthophotos below a specific threshold such as 7.

[cmosig]

Mean for my own flight in Auwald was 9. For a Madeira image from Teja it was 23. For one recently with 85/75 mean was 13.