Human Image Understanding
Abstract
This report summarizes the major research accomplishments performed under AFOSR Grant 99-0231, HUMAN IMAGE UNDERSTANDING. An extensive series of experiments assessing the visual priming of briefly presented images indicate that the visual representation that mediates real-time object recognition specifies neither the image edges or vertices nor an overall model of the object but an arrangement of simple volumes (or geons) corresponding to the object's parts. This representation can be activated with no loss in efficiency when the image is projected onto the retina at another position, size, or orientation in depth from when originally viewed. Consideration of these invariances suggests a computational basis for the evolution of two extrastriate visual systems, one for recognition and the other subserving motor interaction. The experiments suggest that it may be possible to assess the functioning of these systems behaviorally, that is, to split the cortex horizontally, through a comparison of performance on naming and episodic memory tasks. We have developed a neural network model (Hummel and Biederman, 1992) that captures the essential characteristics of human object recognition performance. The model takes a line drawing of an object as input and generates a structural description which is then used for object classification. The model's capacity for structural description derives from its solution to the dynamic binding problem of neural networks: Independent units representing an object's parts (in terms of their shape attributes and interrelations) are bound temporarily when those attributes occur in conjunction in the systems input.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 18, 1991
- Accession Number
- ADA247048
Entities
People
- Irving Biederman
Organizations
- University of Southern California