Tarr, M. J. (1994). Visual representation: From features to objects. in V. S. Ramachandran (Ed.) Encyclopedia of Human Behavior. San Diego, CA: Academic Press, Vol. 4, 503--512.

@InCollection{Tarr1994,
  author = 	 {Tarr, Michael J.},
  title = 	 {Visual representation: From features to objects},
  booktitle = 	 {Encyclopedia of Human Behavior},
  pages = 	 {503--512},
  publisher =    {Academic Press},
  year = 	 {1994},
  editor = 	 {Ramachandran, V. S.},
  volume = 	 {4},
  number = 	 {},
  series = 	 {},
  type = 	 {},
  chapter = 	 {},
  address = 	 {San Diego, CA},
  edition = 	 {},
  month = 	 {},
  note = 	 {},
  annote = 	 {}
}

Author of the summary: Jim Davies, 2006, jim@jimdavies.org

Cite this paper for:

• One way to determine the parts of an object is to look for inward-facing curvature.
• non-accidental properties explaining gestalt effects
• People have both Structural and View-based descriptions

• proximity (things close together apppear to be part of the same group)
• similarity (things that look similar appear to be part of the same group)
• good continuation (collinear edges or boundaries are grouped)

One way to determine the parts of an object is to look for inward-facing curvature.

Perceptual systems are often thought to have biases for non-accidental things, that correspond to the way our world tends to be. This provides an explanation of gestalt effects: close objects tend to be attached, as do similar objects. Surfaces tend to be continuous over large areas, leading to good continuation biases. The idea is that non-accidental properties may form the basis for structure and part recovery. [6]

Object representation can vary on several aspects:

• Frames of Reference. Has implications for how reasoners compensate for orientation discrepancies
  • object-centered (object has a front. Orientation independent.)
  • viewer-centered (relative to eyeball)
  • environment-centered
• Dimensionality. Evidence for both. [7]
  • 3D
  • 2D
• Parts. How to know an object is composed of parts.
  • volumetric primitives. (3d, capture approximate shape, course coding and restricted number of primitives, for finding similarity. unclear whether they can capture all of human visual perception.)
  • locating boundaries (where one part ends and another begins.
• Spatial Relations. How parts relate spatially.
  • implicit (bitmap)
  • explicit (a right of b)
  Another issue is whether the parts are represented with respect to a single point or through a hierarchy (e.g. arm -> forearm -> hand -> finger). [8]
  • single point
  • part hierarchy (benefits: potentially simple top level and implicitly have knowledge of smaller sizes)

• Structural Descriptions tend to be object-centered, composed of volumetric primitives, 3D, explicit categorical spatial relations between parts, hierarchical description. As a result tend to be invariant over changes in size, orientation, position, and perhaps mirror images. Some structural descriptions use volumetric primitives from a restricted class to afford similarity judgments.
• View-based descriptions generally are viewer-centered, boundary locating principle to define parts, 2D plus depth, implicityly encoded spatial relations. Identification tasks realted to viewpoint are often used as evidence for this kind of rep. [9] A variant is "multiple views." Alignment is necessary for similarity judgments. Some mechanism must be in place to know which views to save. One idea is the ones most typically seen.

Summary author's notes:

• page numbers are from the printed version.