Many pictures taken from within a car has some or a lo of the dash board visible. Perhaps it would be possible to detect this and crop the series? Of cause the dash board changes between pictures, when the light changes, but it is usually very little between individual images. The change is more when you turn or the Sun changes.
If you get it working well enough, it would be nice to run it on the existing images. I have several times let some dash board show purposely, because the clouds in the sky would make the auto exposure let the images become way too dark.
But a warning would be good too. I have tried hard to avoid dash board and still gotten it several times.
The main issue I get is not the dashboard itself, but its reflection in the windshield, as @de_vries pointed out. Oftentimes no amount of repositioning gets me rid of it without losing large portions of the scene. It would be great to have a system for removing not only the dashboards, but their reflections as well (and while we’re at it, raindrops and other kinds of mostly static obstructions too).
I just browsed this sequence: http://www.mapillary.com/map/im/DJ7J4nAdEtN2SwSxBD7g7Q/photo
and thought that removal of animals and drops of water on a wind shield would be great. Mapillary gets a lot of images that would never hit a service like Google Streetview, but some of the errors can probably be fixed.
I know you computer vision guys don’t do magic, but you can make it look like that and that is good enough.
Visual image search. Mark a region in an image and find other images with the same/similar object across all of Mapillary.
Sounds like an expensive operation though…
Automatic object tagging (with user feedback and suggestions for objects).
This would be wonderful. I’d probable ‘teach’ it to recognize our local bus stop signs and shelters. Or those signs that indicate there is an underground hydrant. Those make good spatial reference points, so I go out of my way to make pictures of them.
Speaking of reference points. Are you still working on creating a DB of ‘recurring objects’? That’s probably a question for a separate thread though.
We are actually looking into both of your suggestions. Visual search is not as expensive as it sounds. Object tagging requires some infrastructure work. We’ll keep you posted on progress.
I’ve already try to correct the pictures’ exposure, but it’s difficult to do it automaticaly. Could computer vision help for this ?
With the classification you compute on all pictures, it could be easier to exclude unuseful features like clouds, sky, to correct the exposure ?