Multi-view Facial Landmark Detection
Type of document
disertační práceAuthor
Uřičář, Michal
Supervisor
Franc, Vojtěch
Field of study
Umělá inteligence a biokybernetikaStudy program
Elektrotechnika a informatikaInstitutions assigning rank
České vysoké učení technické v Praze. Fakulta elektrotechnická. Katedra kybernetikyMetadata
Show full item recordAbstract
In this thesis, we tackle the problem of designing a multi-view facial landmark detector
which is robust and works in real-time on low-end hardware. Our landmark detector
is an instance of the structured output classi ers describing the face by a mixture
of tree based Deformable Part Models (DPM). We propose to learn parameters of
the detector by the Structured Output Support Vector Machine algorithm which, in
contrast to existing methods, directly optimizes a loss function closely related to the
standard evaluation metrics used in landmark detection. We also propose a novel
two-stage approach to learn the multi-view landmark detectors, which provides better
localization accuracy and signi cantly reduces the overall learning time. We propose
several speedups that enable to use the globally optimal prediction strategy based on
the dynamic programming in real time even for dense landmark sets. The empirical
evaluation shows that the proposed detector is competitive with the current state-ofthe-
art both regarding the accuracy and speed.
We also propose two improvements of the Bundle Method for Regularized Risk Minimization
(BMRM) algorithm which is among the most popular batch solvers used
in structured output learning. First, we propose to augment the objective function
by a quadratic prox-center whose strength is controlled by a novel adaptive strategy
preventing zig-zag behavior in the cases when the genuine regularization term is weak.
Second, we propose to speed up convergence by using multiple cutting plane models
which better approximate the objective function with minimal increase in the computational
cost. Experimental evaluation shows that the new BMRM algorithm which
uses both improvements speeds up learning up to an order of magnitude on standard
computer vision benchmarks, and 3 to 4 times when applied to the learning of the
DPM based landmark detector.
v
Collections
- Disertační práce - 13000 [713]
The following license files are associated with this item: