Mathematics of Sensing, Exploitation, and Execution (MSEE) Hierarchical Representations for the Evaluation of Sensed Data
Abstract
The primary goal of this project was to build fully generative hierarchical scene models and accompanying algorithms and software for inference from still imagery. A secondary goal was to develop a feasible approach to learning these scene models from data. Other goals were less central, but included making connections and contributing to theories of the mammalian visual system, and exploiting descriptive text that may accompany a still image for improved inference. The focus of the Brown team was on single images of street scenes; there was no intention to work with frame sequences. The MSEE goals were ambitious, as were ours. Certainly we failed to meet them and, in fact, our four-or-so year effort can be described as an expedition with a continuously narrowing objective. At the same time, we would suggest that a critical piece of a structure that can support scalable human-level performance has been put in place, new and useful computational tools were discovered, and a new approach to testing vision systems, that places relationships and attributes at the same level of importance as identification, was developed.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 01, 2016
- Accession Number
- AD1010204
Entities
People
- Stuart Geman
Organizations
- Brown University