Comparison to Pt Grey

We are often asked about the difference between the SVS and Pt Grey's stereo processing.  We try to respond as best we can, given our limited knowledge of their system.  Fortunately, CMU's Robotics Institute recently completed a side-by-side competition for a stereo system on their Hyperion lunar rover.  The SVS was chosen; the main points of the report were:

  • Similar stereo range results, with slightly better stereo quality from Pt Grey's trinocular stereo
  • SVS better than 6 times faster at stereo processing

Why is the SVS so much faster?  It uses the same Sum of Absolute (SAD) correlation algorithm as Pt Grey, but SRI has developed a very small memory footprint algorithm (patent pending) that runs in cache.

You can read the full text of the report [PDF, 220 KB].

Feature Comparison Chart

Feature

Calibration

 Rectification Correlation Speed
320x240, 32 disparities
Intel Celeron 600 MHz		 Sub-pixel
interpolation		 Post-filter 3D transform
SVS User calibration
Any binocular stereo hardware		 Yes SAD  59.3 ms / frame
16.8 frames / second		 1/16 pixel Texture, Left/Right Check Yes
Pt Grey Factory set
Pt Grey trinocular hardware only		 Yes SAD 394.7 ms / frame
2.5 frames / second		 1/256 pixel Texture,
Peak Threshold		 Yes

Notes:

  1. Sub-pixel interpolation: 1/16 pixel is more than adequate; interpolation to 1/256 pixel just adds noise bits.
  2. Post-filter: in our experience, the peak threshold is not very effective.  A left/right consistency check, while more expensive, eliminates sporadic close range readings at depth discontinuities.