Mapillary_tools tags wrong time when processing video

I have shot some 2k videos I would like to import into Mapillary. I use the latest version of mapillary_tools on Linux and the command
mapillary_tools sample_video --import_path “$imageDir” --video_import_path “.” --video_sample_interval 0.3 --advanced
where $imageDir points to the directory to store the images in.

At extracted frame number 19 I hold a phone showing GPS time - just in case. The image shows 15:49:15. But the exifdata of the frame shows
Date/Time Original : 2020:07:30 15:32:36.040

This could of cause be the video file that is timestamped wrong, but the exif of the video file says
Date/Time Original : 2020:07:30 15:50:06
This means the time of the video file is a bit off, but at the frame 6 seconds into the video it has not drifted almost 20 minutes.

Has anybody experienced that mapillary_tools in some way will tag the wrong time?

In the 3 cases I have examined, the difference is all 17:31 or 17:33 minutes:seconds. The length of all videos are 17:34. With several rounding errors* there could be something there. (* like I read the GPS time from the actual image but the camera clock might not tick until the next frame = 1 second there)

I have seen errors like this but for much shorter times. In my case 2 seconds was not unusual but I also got into problems when the camera cpu was loaded and dropped maybe 1 B frame in 10. As a result the frame rate was “variable” and skewed the sequence by up to 6 seconds… (1 minute movies)

I did look at the python code in depth one day and found that ffprobe wasn’t being used to get the video start time. I did find though that for BlackVue it was extracting that from the mp4 file raw. I stopped searching at this point.

The video length vs error seems to be the key here. I would guess that the tools are looking at file modified instead of creation time. I vaguely remember that this can also be file-system related, FAT32 for example not supporting this.