Event Representation in Humans and Machines
Abstract
One of the most compute intensive tasks in analyzing naturalistic video is tracking objects and people. Tracking complete databases containing hundreds of thousands of hours of video has traditionally been extremely time consuming and/or expensive. The massively parallel, low arithmetic precision, SIMD architecture proposed by Bates was studied to determine whether it could bring great efficiency benefits to tracking. The slowest subtasks in the tracking pipeline were studied, and it appears that tracking is a task that maps well to the proposed hardware, with the potential for thousands of times speedup, lower energy use, and cost, compared to traditional CPU-based methods.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 30, 2011
- Accession Number
- ADA550111
Entities
People
- Deb Roy
- Joseph Bates
Organizations
- Massachusetts Institute of Technology