Roy, Debaditya and C, Krishna Mohan
(2018)
REPRESENTATION LEARNING FOR ACTION RECOGNITION.
PhD thesis, Indian Institute of Technology Hyderabad.
Abstract
The objective of this research work is to develop discriminative representations for human
actions. The motivation stems from the fact that there are many issues encountered while
capturing actions in videos like intra-action variations (due to actors, viewpoints, and duration),
inter-action similarity, background motion, and occlusion of actors. Hence, obtaining
a representation which can address all the variations in the same action while maintaining
discrimination with other actions is a challenging task. In literature, actions have been represented
either using either low-level or high-level features. Low-level features describe
the motion and appearance in small spatio-temporal volumes extracted from a video. Due
to the limited space-time volume used for extracting low-level features, they are not able
to account for viewpoint and actor variations or variable length actions. On the other hand,
high-level features handle variations in actors, viewpoints, and duration but the resulting
representation is often high-dimensional which introduces the curse of dimensionality. In
this thesis, we propose new representations for describing actions by combining the advantages
of both low-level and high-level features. Specifically, we investigate various linear
and non-linear decomposition techniques to extract meaningful attributes in both high-level
and low-level features. In the first approach, the sparsity of high-level feature descriptors is leveraged to build
action-specific dictionaries. Each dictionary retains only the discriminative information
for a particular action and hence reduces inter-action similarity. Then, a sparsity-based
classification method is proposed to classify the low-rank representation of clips obtained
using these dictionaries. We show that this representation based on dictionary learning improves
the classification performance across actions. Also, a few of the actions consist of
rapid body deformations that hinder the extraction of local features from body movements.
Hence, we propose to use a dictionary which is trained on convolutional neural network
(CNN) features of the human body in various poses to reliably identify actors from the
background. Particularly, we demonstrate the efficacy of sparse representation in the identification
of the human body under rapid and substantial deformation.
In the first two approaches, sparsity-based representation is developed to improve discriminability
using class-specific dictionaries that utilize action labels. However, developing
an unsupervised representation of actions is more beneficial as it can be used to both
recognize similar actions and localize actions. We propose to exploit inter-action similarity
to train a universal attribute model (UAM) in order to learn action attributes (common and
distinct) implicitly across all the actions. Using maximum aposteriori (MAP) adaptation,
a high-dimensional super action-vector (SAV) for each clip is extracted. As this SAV contains
redundant attributes of all other actions, we use factor analysis to extract a novel lowvi
dimensional action-vector representation for each clip. Action-vectors are shown to suppress
background motion and highlight actions of interest in both trimmed and untrimmed
clips that contributes to action recognition without the help of any classifiers.
It is observed during our experiments that action-vector cannot effectively discriminate
between actions which are visually similar to each other. Hence, we subject action-vectors
to supervised linear embedding using linear discriminant analysis (LDA) and probabilistic
LDA (PLDA) to enforce discrimination. Particularly, we show that leveraging complimentary
information across action-vectors using different local features followed by discriminative
embedding provides the best classification performance. Further, we explore
non-linear embedding of action-vectors using Siamese networks especially for fine-grained
action recognition. A visualization of the hidden layer output in Siamese networks shows
its ability to effectively separate visually similar actions. This leads to better classification
performance than linear embedding on fine-grained action recognition.
All of the above approaches are presented on large unconstrained datasets with hundreds
of examples per action. However, actions in surveillance videos like snatch thefts are
difficult to model because of the diverse variety of scenarios in which they occur and very
few labeled examples. Hence, we propose to utilize the universal attribute model (UAM)
trained on large action datasets to represent such actions. Specifically, we show that there
are similarities between certain actions in the large datasets with snatch thefts which help
in extracting a representation for snatch thefts using the attributes from the UAM. This
representation is shown to be effective in distinguishing snatch thefts from regular actions
with high accuracy.In summary, this thesis proposes both supervised and unsupervised approaches for representing
actions which provide better discrimination than existing representations. The
first approach presents a dictionary learning based sparse representation for effective discrimination
of actions. Also, we propose a sparse representation for the human body based
on dictionaries in order to recognize actions with rapid body deformations. In the next
approach, a low-dimensional representation called action-vector for unsupervised action
recognition is presented. Further, linear and non-linear embedding of action-vectors is
proposed for addressing inter-action similarity and fine-grained action recognition, respectively.
Finally, we propose a representation for locating snatch thefts among thousands of
regular interactions in surveillance videos.
Actions (login required)
 |
View Item |
Altmetric
Altmetric