Classifying images

I have thought about this for a while, so I’ll throw it out and please comment.

Sooner or later Mapillary would have to make a feature to classify images. Some areas are so saturated that navigating is very slow (I have contributed to this myself) and some series are pretty bad but better than nothing (I have also contributed to this) but better series may have been added later (I too have done this).

In short: Using all images does not always give a good result and the selection should not be random if possible.

So how to do it?

Imho there should at least be a classification that only looks at technical quality of the image. This allows for the same measurement to be used everywhere to select the best of the possible next pictures.

What do I mean by technical quality?

  • Is the image well exposed - not too dark or light?
  • Is the image sharp? Not too blurry due to e.g. bad focus or movements.
  • Is the image clear of unwanted objects like dust, rain drops, bugs on the window, reflexes in the window, my fingers or the dashboard?
  • Are the colors ok?
  • Is the image straight, i.e. not tilted because a bike is turning?
  • Is the resolution reasonable? One of my action cams only does about 2 mega pixels, which is not that much.

Personally I would grade each image from 0 to 5 (stars?), with 5 being the best. A 5 is a technically perfect image with a sharpness that is what you can expect form a phone or action cam.

4 Likes

@tryl, Good idea! We have been thinking the image quality classification internally as well. Some of the classification tasks you suggested are relatively easier (well-exposed, sharpness). Overall, it is not easy for the computer to achieve good accuracy for all these tasks. We are thinking about integrating user feedback together with auto-classification in image quality evaluation. It is definitely in our roadmap to improve the navigation.

1 Like

I am having the same considerations, after taking something like 50 000 photos in just the city center.

However, I want to add an extra complicating factor. When using Mapillary photos to add information to OpenStreetMap, I am sometimes able to get useful information from otherwise horrible photos. For example a shop name on a window, a half hidden traffic sign, etc.

I am agreeing with a system to present the best photos for a 3-dimensional rendering of the street, but it would be really valuable to have access to even the horrible ones. :smile:

2 Likes

It would be to a great help if you could do some classification or/and grouping of the pictures, It is a almost hopeless task to sort the pictures from Panoramio as many of them have lost their geotag when I imported them.

And I have to do it again, I got “log of’ed” when I was sorting my pictures, Not something that maked my happy:-(

/Erich

For me the distinction between ‘this is a sequence’ and ‘this is a single photo that is not part of a sequence’ would help (as my photos belong to the second group).

This would also help me as a filter once there is an easy way to see photos on a map: I’d click ‘exclude sequences’.

1 Like