We’ve been experimenting with some basic detection methods to blur out sensitive information in our photos. For example cars and people.
The solution is not particularly elegant, but it works pretty well. The blurring logic covers entire cars and people as opposed to faces and license plates.
The biggest issue right now is equirectangular projections. Detection of objects works well for less distorted areas (center of photo) but objects in ‘stretched’ areas around the top and bottom of the image are regularly missed (as the models I’m using are designed for flat images).
Ideally I’d like to get it to a level comparable to Google Street View or Mapillary where less of the image is spoilt (e.g. not an entire car blurred), but identifiable info is consistently removed.
Does anyone have any links to work others working on this problem? Or suggestions?