GoPro Fusion - what is the right offset?


#1

I couldn’t resist getting the GoPro Fusion and cut down the number of trips (to capture both ways) or cut down the effort of setting up one GoPro to the front, one in the back, and optionally one iPhone on the side window.

The resolution is lower than a dedicated device pointing e.g. forward… but it’s 360º!

I’ve setup my device to point sideways, to cut the airflow. The display points to the left side, so I can look up and see if it didn’t crash, so each camera points to the sides. The stitching line, at the front and back, are thus a bit blurry, but that’s ok, the signs are on the side of the road anyways.

I’ll provide further details on my process (it’s a bit painful), including how to get the GPS data correct, but for now I need your help to understand how to setup the offset of my photos, so the sequence points to the front of the car.

Here’s the test sequence I’ve uploaded:
https://www.mapillary.com/app/user/davipt?lat=38.728380299963646&lng=-9.214602100006232&z=17&pKey=wC0QmQewYHhDa3bJ-LvVTQ

I’ll probably delete this sequence and re-upload it with the full sequence, maybe even on a separate account, so I can easily distinguish between the pure 360º routes I’ve done and need to “complete the map”, vs. the previous captures.

As you can see when opening the link, the initial image points to the right of the car.


#2

Hi!
We just bought a GoPro Fusion too. I can’t really help you yet, but please do share your solutions and experiences, I will get back to the thread as soon as we can start testing.


#3

I’m still trying to figure everything out, as Mapillary has been quite slow processing fotos in general, and even slower processing 360, but I’ll try to leave some highlight points here already:

  • update the software. Currently 1.8 on the Fusion and 1.2 on the Mac app
  • the fusion will give up writing gps coordinates when overheating (I assume). Always keep a backup - either the phone with a GPX tracker, or another regular GoPro on the front glass for example.
  • first pass is to stitch the photos. Version 1.2 no longer truncates/rounds the coordinates (they will still be different by a couple of metres, but that’s better than before)
  • Mapillary won’t recognise the GPS from the stitched images.
  • maybe using exiftools to set all dates to be the GPS date, maybe this could solve the exif issue, as the tool shows a warning about exit ordering, and fixes it. I haven’t tried it.
  • however, to make sure all dates necessary for Mapillary are the original ones, and not the stitched ones, I use the ExifTool to re-set them:
    exiftool “-alldates<Composite:GPSDateTime” .
  • additionally I use Lightroom to mask myself and tweak a bit the photos, so the export of Lightroom also fixes the GPS exif.

Other things I’ve learned since:

  • sequences with more than 10K photos… the web only normalizes 10K. I’m not sure if I can normalize the remaining ones.
  • 360 photos takes much longer to be processed than 2D ones. I have sequences of 360 + Phone to the front (was covering a whole block and using the app to know what was missing), plus the GoPro6 to the side (to capture the street names and signs), and those took some days, whilst the 360… well, it makes sense that it’s much harder to process.
  • the normalisation seems to be the same as the others - just point forward, even if the camera is mounted sideways.
  • get a good machine with a good GPU. The Intel Iris Pro (MacBook Pro 2015) is the bare minimum, according to GoPro, but in practice it takes ages, and the whole macOS UI lags for seconds (the software takes the whole GPU cycles and memory).

#4

Hello ! I’m using exiftool to remove the XMP metadata (which prevent mapillary from reading the GPS data) :wink:


#5

Having to remove xmp tags to let mapillary scrape exif seems like a bug. Having to keep a backup gps for a £500 camera also seems like a joke.
I got one cheaper, but reading this now not sure if it was a good choice since I already plan on a multi-hero setup around the car anyway.
Is it better to point the camera to the front since it normalises as forward-facing?