Attention and Awareness Christof Koch

1
Attention and AwarenessChristof Koch
www.klab.caltech.edu
2

3
(No Transcript)
4
How Is Perceptual Saliency Computed?

• Bottom-Up Component very rapid (lt 0.15 sec),
  automatic, task-independent and low-level
  saliency, intrinsic to image.
• Top-Down Component takes longer, depends on
  task, voluntary, coupled to high-level cues
  (faces, sex, dominance in hierarchy).
• Our research is primarily concerned with
  bottom-up image saliency. It will therefore be
  primarily useful when rapidly scanning images
  (e.g. browsing a magazine, video, walking past a
  sales display, network of sensors).

5
Systems Architecture

• Single saliency map encodes bottom-up
  conspicuity and is feed by multiple (42) feature
  maps
• After detecting most salient location, its
  saliency is inhibited (gt Inhibition-of-Return)

6
(No Transcript)
7
Our algorithm (ca. 40,000 lines C code)
evaluates image streams in 42 topographic maps
and a few million neurons representing the
image in addition to a saliency
map. Interesting locations in this map are
sequentially scanned in the order of their
saliency.
8
(No Transcript)
9
(No Transcript)
10
(No Transcript)
11
(No Transcript)
12
(No Transcript)
13
Saliency for Low-Bandwidth Sensor Network

• Network of cheap cameras and microphones
• Only the most salient location/spectral
  components in each sensor are signaled to
  supervisor

14
Conclusions

• A single saliency-map is an efficient way to
  control where to look next
• The real computational issue is normalization
  within and across maps
• Computational architecture based on primate
  visual system yields remarkable machine vision
  performance
• Plethora of commercial and military applications