Week 4

Query-based video
summarization
UCF REU 2016
Jacob Laurel
Graduate Student Mentor: Aidean Sharghi
Professor: Dr. Boqing Gong
Overview
Create fully summarized data set for
egocentric videos for nearly 50 queries,
since no such data set exists
 Implement Baseline detection algorithm
on dataset that uses a Determinantal
Point Process to automatically generate
diverse summaries

Weekly Progress
Selected the annotations to be used by
determining most popular search queries
and concepts on Youtube and Vine
 Continued running DPP experiments,
results are promising
 Added final improvements to UI to allow
for more accurate video annotations and
summarizations
 Uploaded annotation/summarization jobs
to Amazon Mechanical Turk

Annotation Selection

A list of the most popular categories,
queries and hashtags from Youtube and
Vine were studied to try and narrow
down a list of ~50 concepts that could be
used to generate a query-based video
summary
Experimental Methods and User
Interface
Upon testing the time taken to finish an
annotation and the quality of annotations,
the User Interface was tweaked to show
5 frames from a shot that would be
annotated together.
 This yielded more accurate summaries
and annotations
