Key frames in loop closure detection

Source: cometlabs

Most common method to get candidate key frames: use a place recognition approach

• approach based on vocab tree
• feature descriptors of candidate key frames are quantised
  • one colour in image below corresponds to one feature descriptor/‘vocabulary’
  • each point is a ‘word’ that belongs to a vocabulary
  • the words can then be counted and put into a frequency histogram
  • the histogram is used to compare similarity of images
  • I think similar images then get filtered out, so we get key frames