Hi guys,
I have the feeling that the object detection on my GoPro Hero 11 is much better (I guess it has to do with the resolution). So I’m wondering if it would make sense to use my Hero 11 as a second camera at the same time when using the Hero Max from Camera Grant.
But maybe that doesn’t make sense due to data + the merging of normal & 360° images in the “Streetview” view.
Maybe someone has some thoughts on this.
Yes, it makes sense to combine the full-circle 360 image with the higher resolution image from any recent Hero, have added a Max behind the Hero 10 which is mounted on the cycling helmet.
The Max provides a lower resolution image (defined in pixels per degree), but adds context and bits which were outside the Hero’s field of vision.
(To view an example, zoom in to the area near Antwerpen in Belgium, filter to show both koninklijke and konink360 - the Max was added circa end 2023, might like to add date filter as well) - here’s a pre-set URL : Mapillary
Please don’t hesitate to ask …
Thanks,
do you use both cameras in time-lapse, video mode or different settings? I would probably go with video mode on the Hero 11 (I have 4 batteries) and time-lapse mode on the Max.
Both cameras are set to time-lapse, one per sec for the Hero, as I noticed a ‘two steps forward / one step back’ effect in the two pics per sec series after uploading to Mapillary which in turn leads to the ‘viewing angle’ indicator pointing the wrong way; that ‘Echternach procession’ effect isn’t visible when looking at the series in JOSM (the Java OSM editor) which I use to remove duplicates while stopped, and to correct location on pics in underpasses / urban canyons.
Video would allow Mpy to extract more than one pic/sec when cycling faster than two or three m/sec, but unfortunately found that location got out of sync - not just on my Hero10, also on the previous Hero9, and on another contributor’s camera.
This was observed both while standing still at a traffic light, and while cycling along at mildly varying speed - as in up and down slight in- and declines.
There also is the lower resolution in the video stills (4k vs 5k3).
Is it not possible to start/stop the time lapse on both cameras at the same time with the gopro remote watch ?
A bit of n00b here. A few questions out of curiosity to better understand things.
a) What do you mean by object detection? Is that something the GoPro uses to figure out where to focus?
b) Why video instead of pics 2 per second?
c) Why the sync? Aren’t they going to be their own separate sequences on Mapillary with the timing driven by the GPS?
Thank you.
I wasn’t sure if it is really necessary but I will try it out.
I have observed something similar too. However, this has probably less to do with resolution but more with the fact that the AI detection model has been trained mostly on linear perspective imagery than on 360° or fish-eye lens imagery in the past. I am not sure what the distribution of projection parameters in the Mapillary dataset may be today but I am pretty sure that linear projection imagery continues to quite significantly outnumber all other projection parameter imagery. The same AI detection model runs all kinds of projection parameters imagery, so you get better detection rates on linear projection images, or HERO fish-eye lens imagery because it is still closer to linear than 360°.
Well, it certainly will not hurt things and if detections is what you are going for in your imagery then go ahead. 
Indeed, the GoPro MAX really suffers from too low of a resolution for 360° imagery (and sensor brightness imbalance, etc). However, you can best improve overall data quality with the GoPro MAX if you 360° capture both sides of the road in one sequence, preferably along the curb. This way, reconstruction will properly align all other potentially already captured imagery on any given road. Plus, “curb riding the GoPro MAX” should also make house numbers and street name signs legible most of the time in its 360° imagery.
The GoPro does not detect anything. Everything happens on Mapillary servers in the cloud after upload. However, the GoPro does visually detect motion (and by gyro but not for image processing) in a quite sophisticated way and uses that data for example for Hypersmooth or tuning in variable bitrate encoding parameters. Note that GoPros do not detect focus because all GoPros are fixed focus by design.
Oh yes, please do not use video if you can. Better go with shorter or shortest time lapse intervals than with video. If you really must use video then please set your camera to its maximum constant bitrate encoding parameters. To be able to do this on a GoPro you will need a GoPro Labs firmware. If ever, video should be treated as a fallback only option for Mapillary.
