Empirical Study on the Tradeoffs of Action Recognition Models for Industry
Name
Snowdon-jsnowdon-meng-eecs-2022-thesis.pdf
Description
Thesis PDF
Size
10.74 MB
Format
Adobe PDF
Checksum (MD5)
3a6d19c3071293da7615393d4d7df968
Author(s)
Snowdon, Jack
Advisor(s)
Oliva, Aude
Martie, Lee
Date Issued
May 2022
Publisher
Massachusetts Institute of Technology
Abstract
Action recognition has attracted intense attention in the last decade. Advances in deep learning and the availability of large-scale video datasets have drastically improved its capabilities, attracting interest from industry with a variety of use cases. My work at the MIT-IBM Watson AI lab presents a survey analyzing the performance of various existing action recognition methods in the context of construction site analysis and workplace safety. The analyzed pretrained action recognition models were developed in the lab, and encompass a range of popular techniques in the field, each with their own strengths. In addition to developing a general pipeline to train these models on novel datasets, an easy-to-follow guide is presented to make model recommendations for both of the construction site and workplace safety tasks. Although it was hard to make definitive recommendations without knowing the specific hardware constraints, the results obtained and the discussion offer insight into what it feasible and effective with current technology in the problem spaces.
MIT Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
Terms of Use
In Copyright - Educational Use Permitted
Copyright MIT
Persistent DSpace Link
