DiamondClassify
The DiamondClassify project has created a group of libraries intended to support the development of products using the MERL object detection and recognition algorithms. The code is largely based on the original MERL detection and recognition libraries, but substantial improvements have been made to code readability and organization to bring the code to product-level status.
Background & Objective: The MERL object detection and face recognition code had grown into a large and unwieldy library, containing both detection and recognition code intended for use in products, and significant amounts of obsolete and experimental research and training code. The primary goal of the DiamondClassify project has been to break this large monolithic library into several smaller, more manageable units by separating the code into libraries intended for use in products, and libraries solely used for research and training purposes.
Technical Discussion: Last year, our initial work on DiamondClassify focused on the product-level portions of the detection and recognition libraries. This year, our effort has focused on the research and training code, which resides in a separate library. First, the training library was ported to use the existing API provided by the DiamondClassify product-level code. Next, the object detection and face recognition training programs were ported to run on the new training library, on a multiprocessor system. Finally, the face recognition code was modified to run on a single-processor system. As part of this project, significant effort was made to improve performance so that running on a single-processor system would be feasible. As a result, face recognition training now runs approximately 70 times faster on a single processor than in the past. Finally, the Bayesian matching face recognition algorithm developed at MRL was ported to run within the DiamondClassify API. This allows the choice of recognizer to be made at run time by simply providing a different recognizer file to any program using the face recognition API.
| Technical Reports: | |
| Bayesian Face Recognition | |
Technology Area: Computer Vision
Modification Date: January 23, 2007