Tracking multiple independent targets: Evidence for a parallel

tracking mechanism

Zenon Pylyshyn and Ron Storm

presented by

Nick Howe

September 30, 1998 Vision Seminar Slide #2

The Picture

• Humans have three “tracking” mechanisms:– Head motion– Eye motion– Focus of attention

September 30, 1998 Vision Seminar Slide #3

Locus of Visual Attention

• Size debatable. (variable?)

• General agreement that there is just one region of attention.

• Motion of region must be continuous and of constant velocity (no saccades).– Estimates of velocity range from 30 deg/s to

250 deg/s, with median at about 50 deg/s.

September 30, 1998 Vision Seminar Slide #4

The Question

• Do people track multiple objects using a serial process, or some sort of parallel one?– Serial process involves scanning objects

sequentially, updating positions one at a time.– Parallel process hypothesizes some

preattentional mechanism which follows multiple objects simultaneously.

September 30, 1998 Vision Seminar Slide #5

More on Parallel Tracking

• Hypothesizes indices (called FINSTs) which “stick” to image features and track them as they move.

• FINSTs act as indices into the location of the feature, allowing attention to be transferred to them if desired.

September 30, 1998 Vision Seminar Slide #6

The First Experiment

• Purpose of experiment:Confirm that subjects can track multiple objects.

• Features:– Eyes fixated in center of screen.– Ten moving + symbols.– Subjects tracked 1 to 5 target symbols.– Asked to respond when target “flashed”.

September 30, 1998 Vision Seminar Slide #7

Results

• Subjects can track multiple objects at once.– 2% error while tracking one target.– 14% error while tracking five targets.– Response time also increases slightly.

• No performance/time tradeoff.

September 30, 1998 Vision Seminar Slide #8

The Second Experiment

• Designed to rule out serial tracking hypothesis.

• Similar design:– Four targets, four distractors (eight total).– Higher velocities.– Targets always had distractor in vicinity.– Random flashes added as possibility.

September 30, 1998 Vision Seminar Slide #9

Results

• Subjects averaged 87% correct responses.

• By comparison:– Serial model predicts only 8% correct.– Serial model with velocities predicts 20%.– Serial model with guessing predicts 45%.– Hybrid tracking predicts up to 78.6%.

• Thus experimenters rule out serial model.

September 30, 1998 Vision Seminar Slide #10

Conclusions

• Some performance degradation with increase in number of tracked objects

=> “resource-limited” parallel tracker.

• Support for FINST model.

September 30, 1998 Vision Seminar Slide #11

The FINST Model

• FINST tracks features in parallel, without attention.

• Flash is recognized as a “pop-out”.

• After flash event, tracked objects are considered serially as potential sources, using FINST as index into object location.

• Predicts observed results.

Sequence Seeking and Counterstreams: A Model for

Bidirectional Information Flow in the Cortex

Shimon Ullman

presented by

Nick Howe

September 30, 1998 Vision Seminar Slide #13

The Problem

• Common search problem:

Find a link between two representations.– Match sensory impressions with memory.– Perform motor action appropriate to a situation.– Many other problems can be cast in this light.

September 30, 1998 Vision Seminar Slide #14

Hypothesis

• Bidirectional search– Each representation transformed in effort to

match other.– Record of transformations as priming “trace”.– When the two search trees overlap, linking path

is established.

September 30, 1998 Vision Seminar Slide #15

The Picture

September 30, 1998 Vision Seminar Slide #16

The Picture, II

B B*

A*A

September 30, 1998 Vision Seminar Slide #17

Express Lines

• Search spaces may still be huge.

• There should be a role for context effects.

=> “Express Lines” are connections where an activation on one search tree can cause activation on the other.

(Also inhibition.)

September 30, 1998 Vision Seminar Slide #18

Context

• Context and other cognitive effects are included as a prior priming of nodes.

• Priming duration is longer than that of trace priming.

September 30, 1998 Vision Seminar Slide #19

Learning

• Some paths may be more successful than others.

• Over time, commonly successful paths become stronger; i.e., they are explored first.

• Paths may be refined over time.

September 30, 1998 Vision Seminar Slide #20

Biological Evidence

• The streams / counterstreams theory is difficult to verify by experiment.

• Some neurological evidence is supportive.

• Connections in V1 could support such a mechanism.

B B*

A*A

September 30, 1998 Vision Seminar Slide #21

Evidence for Priming

• Several researchers have proposed mechanisms for the type of priming described.

• Whether these mechanisms play such a role has not been established.

September 30, 1998 Vision Seminar Slide #22

Conclusion

• A cute theory, but difficult to verify.

• Complexity?

• Should computer vision attempt to emulate?