[gdal-dev] Potential VRT bug

[Even Rouault]

> I came across some strange behavior while panning around a VRT in QGIS. The
> VRT consists of several large overlapping TIFs where some TIFs have NODATA
> in areas where other TIFs have data. The behavior is hard to explain, but
> it looks like the data read from the files is corrupted. Panning around,
> zooming in and out, anything that triggers another read seems to change the
> way the data looks. It changes colors, gets skewed, etc. The TIFs
> themselves are fine on their own.
 
> I have reduced it down to a small subset of the data which is available here
> (about 2 MB):
 https://s3.amazonaws.com/pschmitt-test-public/data.tar.bz2
> 
> It consists of two TIFs:
> 1.tif - fully valid
> 2.tif - some valid data and some NODATA
> 
> I also included three VRTs which I created with these commands:
> gdalbuildvrt -r near near.vrt 1.tif 2.tif
> gdalbuildvrt -r bilinear bilinear.vrt 1.tif 2.tif
> gdalbuildvrt -r bilinear bilinear-reversed.vrt 2.tif 1.tif
> 
> If I run this command repeatedly, the min/max of the 4 bands always come out
> the same:
 gdal_translate -projwin 488135 4405500 488374 4405348 near.vrt
> /vsistdout/ | gdalinfo -mm /vsistdin/ 
> Run this same command on bilinear.vrt, and the min/max change each time. So
> that's weird.

For the record, the above described issue was found to be fixed per
https://github.com/OSGeo/gdal/pull/1943 (available in 2.4.3)

Even

-- 
Spatialys - Geospatial professional services
http://www.spatialys.com