Pouring massive amounts of images into the 3D reconstruction

Recently I started experimenting with a rig with 4 cameras pointing in different directions. Walking the same route 2 times - one in each direction - gives in theory 8 images for a good panorama.

Well, my comfiguration uses a phone, an action cam with GPS and two action cams without GPS that gets geotagged from another phone. The result looks like this: http://www.mapillary.com/map/im/AQRUQEHafKRpOV3xDFSeNg/photo (press Photo Details and enable Show camera frames).

Some times it is awesome, because there are images of every direction. But some times it seems that Mapillary have a hard time correcting the different GPS signals, that are not exactly overlaid.

Actually I have just replaced the phone with a Garmin Virb Elite, and I might do some work to geotag everything from the same source. Getting different time formats and zones from various devices to match is not easy, though.

Have others tried something simmilar out? Examples?

On my bicycle I sometimes mount two cameras, my new Drift HD ghost as front camera and my old Sony HDR AS20 on the rear rack. Both have a wide angle (170 degrees) so I could capture almost everything. Both don’t have a GPS so I have to add GPS coordinates from my Garmin. I add this with Javawa FotoGeoTag. This is good because both photos have the same location. Problem is that my Sony could only take pictures each 5s (see other thread) and the image quality isn’t that good as the Drift (2MP). I usually set the Drift on 3s per photo so both cameras unfortunately don’t take a picture at the exact same time.

Drift camera in forward direction (aspect ratio is 4:3)

Sony in the backward direction takes 16:9 photos

I wanted to work in similar solution, but when I started i found your problem. It’s so hard join all photos in the same times. not all cameras works and save the time as same way. by example, my in go pro i can set only hour:min but not seconds and when you power on the camera always start at second 0. … that’s a problem for sync multiple cameras. ¿how set many go pro cameras for same time?

You can solver this problem with a very good GPS Precision, by example taking points every 0.5s and finding the correct offset for cameras.

so, is complicated.

I thinking build some arduino/raspberry device with multiples cameras and a gps for solve thats problems. but my funds are empty now :(.

You can set the time on a GoPro (with wifi) if you use the app. The time will automatically adjust to internettime. If you do not use the app, the time will go wrong a bit and it adds upp to many seconds over time (add least my GoPro does)

That is a cool experiment @tryl !

Adding lots of images of the same space is a good plan, and it should make the navigation better. When it does not, it is our fault; it will keep improving.

There are many challenges to solve for this to work better. In the perfect scenario, the cameras would be taken the images at the exact same time and Mapillary would know that they were attached together.

On the capture side, it is tricky to get all of the cameras to take the pictures at the same time. On the server side, we offer no way yet to communicate that a set of sequences were taken simultaneously. This might be possible in the future.

Even with the challenges, keep up with the experiments and let people know what works best.


That sounds great @pau.

Making action cameras take pictures in sync is pretty much impossible. Perhaps it is possible to hack something with my two Xiaomi Yi, but with a Garmin and a Sony in the mix, forget it.

I will try to geotag them from the same GPS, which will improve alignment a bit.

But in general, I feed you a huge load of images with GPS data of varying quality and lets you put it together magically using image recognition. I know you cannot do magic, but some times it may look like that :smile: Keep up the good work!