Monocular depth perception

Depth perception in real life

• In nature, prey animals typically have eyes on either side of their head to maximise field of view, while most predators have forward-facing eyes with overlapping fields of vision (binocular vision) for maximum depth perception. Humans also have binocular vision. (Some exceptions: fruit bats, killer whales)

• We perceive depth, or distance to the objects that we see, based on several visual cues.

  • One of the cues is the parallax  in the two overlapping fields of vision, or the ‘binocular disparity’. ( Science Focus )
  • However, even if vision is impaired in one eye, a human can still perceive depth by using non-binocular visual cues.
  • In fact, according to ( Chen 2018 ), “a typical human uses 14 visual cues to perceive depth, only 3/14 are binocular vision related.”

• Some of the monocular cues that the brain uses:  Science Focus

  1. We know the real size of things
  2. Using perspective, e.g. parallel lines converging to a perspective point
  3. Motion parallax

Depth perception in computer vision

• Depth is not recoverable from a single camera image.
• This is called the “scale availability issue.” s.  monocular cameras , whereby the scale is the factor that relates the estimated camera position to the real camera position
• To work around this, there are several options, such as using stereo or RGBD cameras.
• However, even if we are restricted to monocular cameras, using similar depth cues such as those used by the human brain in monocular vision can help recover scale.
  • Some algorithms compare the obtained visual measurements to an object of known scale, e.g. using a calibration checkerboard
  • Other algorithms make use of motion parallax , e.g. the  DefSLAM initialisation