OpenTrailView - more imagery of Italy available, plus weekend 19/20 Dec shutdown announcement

Publishing Platforms, Workflows and Tools
#1

Hi,

Just a quick update on OpenTrailView (https://opentrailview.org) - there are now around 2000-2500 panos of the Lake Como area, Italy, available on the platform plus a smaller number in a nearby area of Switzerland.

These were contributed earlier in the year but only now have I run them all through face/license-plate blurring software (understand.ai anonymizer and github.com/tyndare/blur-persons; see later reply). I haven’t checked all of them manually, but to be quite honest that is impracticable, and there is a “report privacy issue” link so I hope I have done “enough” to address privacy concerns.

One evidence of the power of 360 imagery is shown, incidentally, to my reaction to these panos. Lake Como looks a very beautiful area and were I to have the time, it would be an area I would like to visit - as a direct result of viewing these panos! I think 360 imagery could be a very powerful tool for tourist organisations to promote visits to their area, and precisely why we need a completely free, and open, panos platform which anyone can deploy on their own server.

Also, a quick announcement about an impending shutdown - OpenTrailView will be unavailable weekend of Dec 19/20. This is because I will be updating the underlying database of European OSM data which is likely to take a while and needs to be done while no-one is accessing the server.

Thanks,
Nick

#2

You blurred out the entire car… isn’t that a bit much? Wouldn’t just blurring out the license plate be enough?

#3

@Eesger This was partly experimenting with different blurring tools. I found the one which worked the best - but which was unfortunately rather slow - was this one which segmented the pano into smaller squares.

This took around 6 mins per pano - too slow really for blurring multiple panos though it is very good.
The understand.ai one worked well in the majority of cases, though I found that it would not blur two photos of a child looking partly away from the camera in a countryside location and it also failed on one license plate that was on the border of being readable in the unblurred pano,

So what you’re seeing here is partly the result of experimentation.

The trouble is that it’s hard to get a definitive answer on what our legal obligations are, without the expensive option of taking legal advice (not easy for a not-for-profit project). The understand.ai anonymizer is usably quick, and works well in the majority of cases, but might fail on a few where the face or plate is not in full view. I would like to use only a fast anonymizer - as I mentioned above, I hope that using a reasonably good one plus a “Report privacy issue with current pano” link will mean that I am making a best-effort to consider privacy and take action.

#4

Hmm, that’s 240 images a day… what system was that? would need more cpu power for that option I think :wink:

Still… would be best… have you played with image resolutions…? I guess smaler images would go faster? The quality would offcourse get less the smaler the image… but it the output must be able to give (relative) position results… so that you could blur the larger image with that info…

I have taken a good look at the system of Mapillary. They detect points to blur and place a circular blur around that point. Some detected points have circles that are only a half dozen pixels in diameter! Sometimes there are a lot! But the basic system… they detect points and do a circular blur that fits the detected area… that’s a good system!

#5

It is recommended to run most machine learning based detection software on the GPU, but I am running it on a cloud server which AFAIK has no GPU. It’s quite a powerful server (16GB RAM) but still running rather slow. It seems to be barely faster on lower-resolution panos than higher-resolution.

Might be worth trying the same software with less segmentation, that might speed things up.

#6

Then, again, the way to go is modular. Store which filter you have used (and keep the original)… one can choose to run the ‘perfect solution’ slowly and the quick one first… ?

Or if someone has an (local) GPU server that can do it, then the ‘perfect solution’ can be chosen…

HQ | Blog | View / Upload Photos | Become a Trekker | Facebook | Instagram