This Grab Tech’s blog post reads very promising. Actually, this is something that many people have been waiting for from Mapillary. Apparently, it takes people who are closer to that manufacturing and market closer. Unfortunately, the blog post does not say anything about availability or price of the camera. Nevertheless, things look impressive enough and read very promising so far. The camera should be even capable of some AI on the edge. However, I do not agree with every assumption and capability made by Grab Tech’s engineers:
Privacy information detection
Face and license plate blurring on the edge (while be it with AI) is a great feature not only for mapping but also for publishing imagery in general in the digital age. However, every AI feature must have an override switch!
Scene recognition model
Image quality (IQ) checking AI model
Object detection AI model
If anything, these things should happen on the backend. A camera should not rate or qualify/disqualify the images you capture. Just think about all the false positives. Furthermore, there is a risk of creating a single purpose camera. Devices in general, and digital devices especially should not patronize users. They can and sometimes should warn users about potential dangers or potentially illegal actions but users (people) should decide what to do and not to do, not machines.
Cameras should not process or filter images too much either, be it with AI or static filters. When capturing a photo we should be interested in the physical truth, not some digital glitter processed output stream. We should want to record — as the name in photography says — photons, not produce some prettified pixels.
Their “quadcam” setup for panoramic images also looks interesting. Full spherical imagery is not always really necessary for many purposes, including mapping. What is often more important then full 360° is ease of use, capture, and stitching.
I am also surprised that they deem a 12 MP sensor as sufficient for street-view mapping, especially from vehicles. My computations have lead me to believe that a 4896×3264, ~16 MP, 3÷2, ~2.45µ per pixel sensor with a linear lens of ~13 mm focal length is probably the ideal candidate combo for this purpose. But hey, this is just my opinion based on my assumptions.
Anyhow, I would be happy if Mapillary would get in touch with Grab Tech and perhaps opt for some collaboration in this space.
KartaCam has a built-in 4G LTE module that ensures it is always connected and can be remotely managed. The KartaCam management portal can monitor camera settings like resolution and capturing intervals, even in edge AI machine learning models. This makes it easy for Grab’s map ops team and drivers to configure their cameras and upload captured images in a timely manner.
This means, this device is fully under control by the grab map ops team. This is clearly not a “tool” to be used by independent, individual mappers, but equipment to be installed on vehicles of contracted companies with low-cost labor fleets, such as ride-hailing and delivery services.
In this setting all the edge AI items you mention also make much more sense.
This means, this device is fully under control by the grab map ops team.
This is clearly not a “tool” to be used by independent, individual mappers…
Right, I must have over read that part. Good catch. However, from an engineering point of view this is mostly a software design option. The LTE module is not a bad idea per se because many contributors want to do all of their capturing on one device only and can be confined to rural areas where cell phone networks are basically the only viable option to accomplish uploads. Sure, no independent contributor would want that sort of oversight/surveillance from a camera vendor. However, that LTE remote control evil only works as far as signal coverage goes.
What I was trying to say was that they seem to be on the right track for creating a purpose built camera for capturing street-level imagery. So far, nobody has made a really faithful attempt at this specific use case, not even Mapillary with the BlackVue.
If gopro or insta360 would make a few less lazy software and design decisions we’d be good to go. Both those 360 cameras “could” be the perfect tool and both are made frustrating by the companies’ indifference to the mapping market.
I’m (not very optimistically) hoping the gopro Max 2 steps things up a notch, but past experience would tell me not to get my hopes up. sigh
Yeah, generic action cams are a good starting point for street‑level capturing and mapping but they still lack a few key things that limit their full potential in this regard, like sensor parameters (size & resolution), manual locking of photographic parameters (ISO speed, focus, exposure time limits, white balance), properly working compass, de‑duplication of images while stopped, and ease of upload. Actioncams are designed and built for — like the name says — action (and YouTubers), not still image photography. Furthermore, an actioncam already serves a niche market, so mapping is even more of a niche market. Hence, I believe nothing will happen in this regard unless mapping platform services step in an work together with camera makers on this specific subject.