CATEGORY LEVEL POSE ESTIMATION VIA INSTANCE CANONICALIZATION
Author(s)
Walker, Tevon
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
This thesis describes a study of pose estimation. Pose estimation is the determination of
an object’s position and orientation in some coordinate frame. Pose estimation has applica-
tion in robotics, augmented/virtual reality, and human-computer interaction, among others.
An overview of the field is given along with an explanation of some key challenges. First,
this work specifically details instance-level pose estimation and its application. Instance-
level pose estimation is the setting in which one specific instance of an object is used when
querying pose. For example, a particular make, model, and year of a car may be chosen, and
an instance-level method will predict the position and orientation of several different poses
of that one particular car. Next, the thesis discusses category-level pose estimation and talks
about some key differences between this and the instance-level setting. In the category-
level setting, a method is responsible for predicting poses of several different instances of a
particular category. Alluding to my earlier example using the car, in this setting, instead of
predicting pose on one particular vehicle, a category-level algorithm will predict pose on
any given object from an entire category, say sedan or pickup truck. This is a much tougher
task, as now the algorithm must not only generalize to the various poses of the object, but it
also must learn the general shape of the category. Although category-level methods handle
different instances within a category better than instance-level methods, the bench-marked
performance of these methods is not as high as their instance-level counterparts in terms of
pose prediction accuracy. Finally, a novel idea for improving category-level pose estima-
tion is presented along with some experiments showing this idea being applied. The idea is
to reduce the category-level problem to a two stage process of transforming any given cate-
gory instance to a canonical instance, which is a generic shape that represents all instances
of a category, and then predicting the pose of the transformed instance. Here, this trans-
formation of any given category instance to canonical instance is deemed canonicalization.
Through some experiments, very slight improvement of pose estimation using canonical-
xiiization as a front-end preprocessing step over using an instance-level method alone applied
in a category-level setting is shown. Although the canonicalzation shows promise, when
used in conjunction with the instance-level method, the improvement is slim.
Sponsor
Date
2022-01-14
Extent
Resource Type
Text
Resource Subtype
Thesis