Machine Learning in Crowdsourcing for Video Annotation

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering

Title: "Machine Learning in Crowdsourcing for Video Annotation"

by

Mr. LEE, Min Hyung

Abstract:

Obtaining annotated video sequence data is necessary for training object
recognition and tracking algorithms, which locate the moving objects and
track their motion in a video. However, due to the complexity of bounding
objects manually inside a video, the cost of obtaining such annotated data
is high. To address the scarcity of labeled data, a crowdsourcing platform
equipped with video annotation capabilities, called VATIC, can be used to
distribute the labeling tasks to the crowd through the Internet and
receive the corresponding annotation results. However, due to high
variability in the domain knowledge and effort among different workers,
the quality of the labels is impaired. We propose two novel methods which
seek to aggregate the annotation results from different workers to give
more accurate result without knowing the ground truth labels. The first
proposed method uses robust estimation to iteratively compute the average
bounding box and weighs the results of different workers according to
their distance from the average. Afterwards, temporal smoothing is applied
to incorporate the temporal dimension of a video. The second proposed
method uses a continuous-state hidden Markov model to probabilistically
model the crowdsourcing setting. Using a generalized Baum-Welch algorithm
to learn the model parameters, this method searches for and returns the
most likely sequence of ground truth labels. Using video annotation data
collected from different automatic object trackers, which simulate the
workers in the crowdsourcing platform, we compare the performance of the
two algorithms empirically.

Date            :       29 April 2014 (Tue)

Time            :       3:00pm to 4:00pm

Venue           :       5561 (lift 27)

Advisor         :       Prof. Dit-Yan YEUNG

2nd Reader      :       Dr. Brian K.W. MAK