Biologically Inspired Circuits for Visual Search and Recognition in Complex Scenes
Abstract
Here we report on the advances and insights derived from studying and developing computational algorithms for object recognition, feature-based attention and visual search. As part of these efforts, we have created and documented software that includes feed-forward combination of selectivity and tolerance in visual recognition feedback signals for attention, search and object completion. We have quantitatively characterized and evaluated the performance of the system under a variety of different recognition problems with varying levels of difficulty, different levels of approximation to real-world recognition problems and different degrees of temporal dynamics. These measurements provide state-ofthe- art benchmarks for different recognition problems. In particular, we evaluated (a) Single objects on uniform backgrounds and transformations of those objects (scale, position, viewpoint, illumination); (b) Combination of multiple objects on uniform backgrounds; (c) Single objects embedded in natural backgrounds; (d) Faces and objects in commercial movies. We have made progress on three main fronts that involve extensions and improvements to the existing software: (i) addition of feedback and recognition of occluded objects; (ii) Initial optimization of radial basis function centers in intermediate processing stages; (iii) Visual search in cluttered scenes.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 01, 2013
- Accession Number
- ADA579012
Entities
People
- Gabriel Kreiman
- Tomaso Poggio
Organizations
- Massachusetts Institute of Technology