First of all, 
points for your efforts! It looks like you have figured it out on first attempt. 
Yes, overlap is required (for best results). However, the amount of overlap area required varies due to the feature point detection algorithm’s random nature (not to be confused with point feature detection, like fire hydrants, benches, street lamps etc) and the number of feature points per image. The number of feature points per image is a 3D reconstruction parameter that heavily impacts runtime complexity and RAM footprint. Mapillary has to carefully tune this parameter, depending on image resolution, runtime, RAM size, and available computing capacity. So, please be lenient with Mapillary should not all of your images combine. From what I have observed, the overlap does not need to be large. Usually, 5% should suffice.
Please be patient for reconstruction do its magic. 
Logically, you are correct. But, reconstruction is unaware that it can make this special case assumption for your imagery. Mapillary’s 3D reconstruction must work on all kinds of imagery and be able to combine images from different sources, if possible. This is why it is best to have images overlap in a sequence and between sequences.
On a final note, the image quality is jaw dropping. 
It definitely makes a great addition to Mapillary’s dataset.
@osmplus_org So much for Is Anything Besides 360° Photography Still Worthwhile These Days?