Correspondence-Free Dictionary Learning for

Cross-View Action Recognition

Fan Zhu

Ling Shao

Department of Electronic and Electrical Engineering

The University of Sheffield
Sheffield, S1 3JD, UK
Email: fan.zhu@sheffield.ac.uk

Department of Electronic and Electrical Engineering

The University of Sheffield
Sheffield, S1 3JD, UK
Email: ling.shao@sheffield.ac.uk

AbstractIn this paper, we propose a novel unsupervised

approach to enabling sparse action representations for the crossview action recognition scenario. Superior to previous cross-view
action recognition methods, neither target view label information
nor correspondence annotations are required in this approach.
Low-level dense trajectory action features are first coded according to their feature space localities within the same view
by projecting each descriptor into its local-coordinate system
under the locality constraints. Actions across each pair of views
are additionally decomposed to sparse linear combinations of
basis atoms, a.k.a., dictionary elements, which are learned to
reconstruct the original data while simultaneously forcing similar
actions to have identical representations in an unsupervised
manner. Consequently, cross-view knowledge is retained through
the learned basis atoms, so that high-level representations of
actions from both views can be considered to possess the same
data distribution and can be directly fed into the classifier. The
proposed approach achieves improved performance compared to
state-of-the-art methods on the multi-view IXMAS data set, and
leads to a new experimental setting that is closer to real-world
applications.

I. I NTRODUCTION
In the past few years, along with the explosion of online
image and video data, computer vision based applications
in image/video retrieval, human-computer interaction, sports
events analysis, .etc are receiving significant attention. Also,
as can be anticipated, future products, such as the Google
Glasses, which can essentially revolutionize traditional humancomputer interaction , will bring more requirements and
challenges to computer vision algorithms. As an important
topic in computer vision, human action recognition plays a key
role in a wide range of applications. Many approaches [21],
[19], [31], [27], [7], [24], [8], [28], [20], [30] are proposed,
however, some challenges still remain in real-world scenarios
due to cluttered background, view point changes, occlusion
and geometric variations of the target.
Recently, novel strategies have been proposed to represent
human actions more discriminatively. These representations
include optical flow patterns [5],[2], 2D shape matching [15],
[26], [12], spatio-temporal interest points [4], [14], trajectorybased representation [17], .etc. Many state-of-the-art action
recognition systems [18], [22], [11] are based on the bag-offeatures (BoF) model, which represents an action video as a
histogram of its local features. When cooperating with infor-

mative low-level features on detected spatio-temporal interest

points or densely sampled 3D blocks, the BoF model and its
variants yield encouraging performance in many challenging
and realistic scenarios [22]. However, such results are achieved
under a fixed viewpoint or within limited view point variations,
i.e., the discriminative capability of such representations tends
to significantly degrade when the view point variations are
increased. Thus, we aim to seek a high-level feature representation that brings action videos captured from different
view points to the same feature space, while keeping its
discriminative power and allowing the data to satisfy the
smoothness assumption (which implies that data points which
are close to each other are more likely to share the same
label.) in supervised learning. Many recent efforts have been
paid towards this direction. One typical line of attack is to
infer the three-dimensional scene structure based on the given
cross-view actions, where the derived features can be adapted
from one view to another utilizing geometric reasoning [6].
Junejo et al. [8] applied a temporal Self-Similarity Matrix
to store distances between different pairs of actions for a
view-invariant representation. In [13], a bipartite graph is
built via unsupervised co-clustering to measure visual-word to
visual word relationship across different views so that a highlevel semantic feature that bridges the semantic gap between
the two Bag-of-Visual-Words (BoVW) vocabularies can be
generated. Li et al. [10] adopted the conceptual idea of virtual
views to represent an action sequence continuously from one
observers viewpoint to another, and similarly, Zhang et al.
[28] utilized a kernel-based method to capture all the virtual
views on the virtual path instead of sampled views to keep
all the visual information on the virtual path and eliminate
the parameter tuning procedure. Zheng et al. [29] adopted
the K-SVD algorithm [1] to construct an over-complete transferable dictionary pair, which encourages actions taken from
different viewpoints to possess the same representation. These
methods require either labeled samples in the target view or
correspondence annotations, which, however, are expensive or
impossible to obtain in many scenarios. Our approach is most
similar to [29], however, there is one significant difference in
terms of the training data requirement between our approach
and the one adopted in [29] that we learn the cross-view
action representation in an unsupervised manner and action

Fig. 1. The flowchart of our framework. Low-level dense trajectories are first coded with LLC to derive a set of coding descriptors. By pooling the peak values
of each dimension of all local coding descriptors, a histogram that captures the local structure of each action is obtained. Dictionary learning is conducted
utilizing randomly selected actions from both views, then source view training actions and target view testing actions are coded with the learned dictionary
pair to obtain the cross-view sparse representations.

correspondences across the source view and the target view

are not required in our learning phase. Such elimination of
the strict training data requirement is very useful and can be
seen as a significant progress in cross-view action recognition
since neither the labeled training data in the target view nor
the correspondences across the source view and the target view
are handy to obtain in most real-world applications.
As an attempt towards real-world applications, our approach
addresses the cross-view action recognition problem utilizing
only labeled source view actions and unlabeled target view
actions. In order to capture the local structure of actions from
each view individually, the dense trajectories features [22] are
first coded by the Locality-constrained Linear Coding (LLC)
[23] layer. The view knowledge transfer is performed by an
efficient dictionary learning method [16], which brings the
query action in the target view into the same feature space
of actions in the source view. The construction flowchart of
the cross-view sparse representation is shown in Fig. 1. This
paper makes the following contributions:
? By captureing both local action structures and the crossview knowledge, the proposed representation guarantees its
discriminative capability over different action categories as
well as different observation viewpoints.
? In accordance with the initial intention of transfer learning,
the proposed approach is unsupervised, and only requires
action labels from the source view.
? We give an in-depth review of dictionary learning and coding
methods under different constraints.
? The experimental results are promising, and they lead to a
new setting towards real-world applications.

II. M ULTI - LAYER C ODING

A. Problem Statement
We define XS = [x1S , , xnS ] Rdn as n d-dimensional
low-level features extracted from source view actions and
dm
XT = [x1T , x2T , ..., xm
as m d-dimensional lowT ] R
level features extracted from target view actions, where XT
are unlabeled and XS are labeled. Based on the fact that action
videos from two different viewpoints must contain the same
action, we assume there exists a high-level action representation shared between the two videos captured from different
viewpoints. Given any XT , we can always find corresponding
actions in XS . The aim is to discover the connections between
XT and XS , and exploit such connections across both views
to infer the functions FS : XS G and FT : XT G, where
G satisfies the smoothness assumption, so that FS and FT can
bring the cross-view actions into the same feature space.
B. Dictionary learning
We start with the general sense of dictionary learning.
Consider a projection matrix D RdN , and the coefficient
0
= {i , 2 , , d0 } Rd N , and suppose that a data sample xi X can be reconstructed by the linear transformation
of the coefficient i through the projection matrix plus the
reconstruction error as:
xi = Di + reconstruction error.

(1)

If we define an objective function:

F(D, i ) = kxi Di k22 ,

(2)

the coefficient i can be estimated by minimizing F(D, i )

subject to appropriate constraints. If d0 d, the solution

to the unconstrained optimization problem in Equation (2)

is not unique, thus it leads to the over-fitting problem. In
order to give more discriminative and representative solutions
when estimating , additional constraints on are necessary.
The commonly used constraints include regularizing with l0 ,
l1 , l2 -norms, where l0 -norm and l1 -norm are also known
as Sparsity-inducing norms and l2 -norm is also known as
Euclidean or Hilbertian norm. In general, l2 -norm has well
developed theory and algorithms, and it has been applied to
non-linear predictors, non-parametric supervised learning and
kernel methods. On the other hand, the developing Sparsityinducing norms attract more attention in recent years, and
applications that benefit from the sparsity include compression,
regularization in inverse problems, etc. The lowest l0 -norm,
which indicates the solution with fewest non-zero entries, is
studied in this work. With such a constraint, Equation (2) can
be formulated as:
F(D, i ) = kxi Di k22 ,

s.t.ki k0 T,

1
kXs Ds s k22
2

1
+ kXt Dt t k22 + ([s , t ])
2
s.t., [kis k, kjt k]0 T,

(6)

where DS and DT are the source view dictionary and target

view dictionary, and s and t are the the coefficients of
the sparse decompositions. () is defined as the smoothness
0
measurement of [s , t ] Rd (n+m) . A small () indicates
the data in the new feature space are smooth, i.e., those data
points close to each other are more likely to share the same
action label, and thus yields a preference for geometrically
simple decision boundaries. To compute (), we define a
transformation matrix Q:
(x1t , x1s )

..
..

.
.

Q=
..

.
1
(xm
t , xs )

..

(x1t , xns )

..

.
,

..

.
n
m
(xt , xs )

(7)

where we apply the Gaussian kernel (xit , xjs ) to conduct a

fuzzy search across both domains:
2

Instead of VQ, we apply LLC to represent the low-level

action features by multiple bases. In addition to achieving
less quantization error, the explicit locality adaptor in LLC
guarantees the local smooth sparsity. Given the low-level
action representation V = [v1 , v2 , ..., vn+m ] Rd1 (n+m) and
the codebook B with M entries, the LLC code is constructed
by minimizing the following function:
kvi Bci k2 + kqi  ci k2 ,

min

s R,t R

C. Locality-constrained Linear Coding

N
X

F(DS , DT ) =

(3)

where T is the sparsity constraint factor that limits the number

of non-zero elements in the sparse codes, so that the number
of items in the decomposition of each i is less than T . In
the case that T = 1, i.e., each i is represented by a single
basis in the codebook, Equation (3) is equivalent to Vector
Quantization (VQ).

FLLC (ci ) =

for each training sample xi , we have to minimize the divergence between Xs and Xt :

(4)

i=1

where  denotes the element-wise multiplication, and qi

RM is the locality adaptor that gives different freedom for
each basis vector with respect to its similarity to the low-level
feature vi . Specifically,


dist(vi , B)
,
(5)
di = exp

where dist(vi , B) denotes a set of Euclidean distances between vi and all the bj , and adjusts the weight decay speed
for the locality adaptor. By subtracting max(dist(vi , B)) from
dist(vi , B), di is normalized to be between (0, 1]. Applying
max pooling on all the local codes according to each dimension of the codebook, a global representation xi is obtained.
D. Cross-view dictionary learning
When considering action samples from both the source view
and the target view, two corresponding dictionaries are learned
simultaneously. In addition to finding a good representation

2
j
xi
1
t xs
(xit , xjs ) = e( 2 ) .
2

(8)

In order to build the one-to-one correspondence across both

views, the maximum element in each column of Q is preserved
and set to 1 while the remaining elements are set to 0:

1, if Q(i, j) = max(Q(:, j))

(9)
Q(i, j) =

0, otherwise.
Thus, Equation (6) can be expanded as:
4

F(DS , DT ) =

min

s R,t R

1
kXs Ds s k22
2

1
+ kXt Dt t k22 + kt Qs k22
2
s.t., [kis k, kjt k]0 T,

(10)

Assuming Q leads to a perfect mapping across XS and XT ,

i.e., each connected pair through Q has an identical representation in the new feature space, i.e. kTt QTs k22 = 0, and
correspondingly kXtT QXsT k22 = 0. Thus Equation (6) can
be rewritten as:
1
4
fn (DS , DT ) = min kXs Ds k22
R 2
1
(11)
+ kXt QT Dt k22 ,
2
s.t., kk0 T,
where represents the set of identical coefficients.

E. Optimization
The problem of minimizing the empirical cost is not convex with respect to DS and DT . We further define X =
= (DsT , DtT )T . Thus, optimizing
(XsT , (Xt QT )T )T and D
Equation (11) is equivalent to optimizing the following equation1 :
1
4
22
) =
fn0 (D,
min kX Dk
R 2

s.t., kk0 T.

(12)

It is not hard to observe that such an optimization problem

can be solved using the K-SVD algorithm [1]. Specifically,
Equation (12) can be solved in an iterative manner through
both the dictionary updating stage and the sparse coding stage.
In the dictionary updating stage, each dictionary element is
updated sequentially to better represent the original data in
both the source domain and the target domain. When pursuing
the sparse codes are frozen, and
a better dictionary D,
each dictionary element is updated through a straightforward
solution which tracks down a rank-one approximation to the
matrix of residuals. Following K-SVD, the k-th element of the
and its corresponding coefficients, i.e., the k-th
dictionary D

row in the coefficient matrix

P , are denoted as dk and k
respectively. Let Ek = X j6=k dj j and we further denote
k as the results we obtain when all zero entries in
k and E
k and Ek are discarded, respectively. Thus, each dictionary
element dk and its corresponding non-zero coefficients k can
be computed by:
k dk k k22 .
< dk , k >= arg min kE

(13)

k
dk ,

G. Classification
We use the multivariate ridge regression model to train a
:
linear classifier W
= arg max kH W k22 + kW k22 ,
W

(15)

where H is the class label matrix of the source view training

data , and W denotes classifier parameters. This yields the
= HX T (T + Z)1 , with Z being an identity
solution W
matrix. For a query target view action video, we first compute
its sparse representation i and then estimate its class label
i , where l represents the label vector. Finally,
vector l = W
its label is the index that corresponds to the largest element
in l.
III. E XPERIMENTS AND R ESULTS

The approximation in Equation (13) is achieved through

performing a Singular Value Decomposition (SVD) operation
k :
on E
k ) = U V T
SV D(E
dk = U (:, 1)

updated by minimizing quadratic functions, and the remaining

dictionary elements are updated upon the previous updates.
Consequently, the MSE of the overall reconstruction error
is monotonically decreasing with respect to the dictionary
updating iterations. In the sparse coding stage, computation of
the best matched coefficients under the l0 -norm constraint
also leads to a reduction in MSE conditioned on the success
of the OMP algorithm. Finally, since MSE is non-negative,
the optimization procedure is monotonically decreasing and
bounded by zero from below, thus the convergence of the
proposed dictionary learning method is guaranteed. The typical
strategy to avoid the optimization procedure getting stuck in
a local minimum is to initialize the dictionary with a few
different random matrices in several runs. Such a strategy is
applied in our approach.

(14)

k = (1, 1)V (1, :),

where U (:, 1) indicates the first column of U and V (1, :)
indicates the first row of V .
In the sparse coding stage, we compute the best matching
projections X of the multidimensional training data onto the
using an appropriate pursuit algorithm.
updated dictionary D

Given the fixed D, the optimization of Equation (13) is still

generally NP-hard under the l0 -norm constraint. Therefore the
Orthogonal Matching Pursuit (OMP) algorithm is adopted to
approximate the solution in a computationally efficient way.
F. Convergence analysis
The convergence proof of the iterative optimization procedure can be given as follows. In the dictionary updating stage,
each dictionary element and its corresponding coefficients are
1 To prevent D
from being arbitrarily large, column-wise l2 -norm is applied
Without such a constraint, all elements within the sparse coefficients
to D.
may possibly become zero.

We evaluate our approach on the IXMAS multi-view action

dataset [25] (exemplar actions are shown in Fig. 2), which
contains eleven action categories, e.g., walk, kick, wave, etc.
Each action is performed three times by ten actors taken from
five different views. We follow the leave-one-action-class-out
scheme [10] to separate the data, where we consider one action
class (called an orphan action) in the target view, and exclude
all the videos of that class when learning the dictionary pair.
Samples used for dictionary learning are randomly selected
from the non-orphan actions. In accordance with previous
work [10], 30% of the non-orphan actions are chosen from
each view separately. The experiments are conducted on any
possible combinations of view pairs, i.e., twenty combinations
in total are considered.
We use the dense trajectories [22] as the primary feature to
represent raw action video sequences. Dense trajectories are
extracted with 8 spatial scales spaced by a factor of 1/ 2,
and feature points are sampled on a grid spaced by 5 pixels
and tracked in each scale, separately. Each point at frame t is
tracked to the next frame t + 1 by median filtering in a dense
optical flow field. To avoid the drifting problem, the length of
trajectory is limited to 15 frames. Additionally, HOGHOF [9]
and Motion Boundary Histogram (MBH) [3] are computed
within a 32 32 15 volume along the dense trajectories,
where each volume is sub-divided into a spatio-temporal grid

Fig. 2. Exemplar frames from the IXMAS multi-view action recognition dataset. The columns show 5 action categories, including check watch, cross arms,
scratch head, sit down, wave, and the rows show all the 5 camera views for each action category.
TABLE I
P ERFORMANCE COMPARISON OF ACTION RECOGNITION WITH AND WITHOUT KNOWLEDGE TRANSFER .
%
Camera
Camera
Camera
Camera
Camera

0
1
2
3
4

Camera 0
woTran
wTran
25.76
92.42
16.06
92.73
12.42
94.24
12.42
93.94

Camera 1
woTran wTran
23.03
92.42
7.27
92.42
9.39
93.33
10.91
93.03

Fig. 3. Performance comparison with state-of-the-art methods.

of size 2 2 3 to impose more structural information in the

representation.
The experimental results are shown in Table 1, where rows

Camera 2
woTran
wTran
23.94
89.09
35.15
90.61
26.36
90.91
10.3
92.12

Camera 3
woTran
wTran
26.67
91.52
33.33
92.42
29.39
92.12
17.27
95.15

Camera 4
woTran
wTran
30.61
90.00
30.30
90.30
34.55
90.91
30.91
90.30
-

correspond to the source views and columns correspond to

the target views. For each view pair, we show both results for
woTran and wTran settings, which denote action recognition with and without knowledge transfer respectively. In
both settings, low-level dense-trajectories are first coded with
LLC in each individual view. The woTran setting is treated
as a normal supervised classification task in the same feature
domain, so that the LLC codes are directly fed into classification. On the other hand, in the wTran setting, the LLC
codes are further decomposed to sparse linear combinations of
basis dictionary elements, which are learned utilizing samples
from both views. We construct a codebook with 1, 000 visual
words for the LLC codes, and the learned dictionary pair
for cross-view knowledge transfer is set to be size 90. The
average accuracies are 22.52% and 92.00% for woTran and
wTran respectively. In all the cases, the proposed method
outperforms those that directly use classifiers trained on the
source view to predict action labels in the target view, where
the most significant improvement is 85.15%. We also compare

the average performance for each camera view of the proposed

approach with state-of-the-arts methods in Fig. 3. Clearly, our
approach significantly outperforms others even though with a
stricter setting.
IV. C ONCLUSION
In this paper, we have presented an unsupervised dictionary
learning method to address the cross-view action recognition
problem. By setting up virtual connections across the source
and target view samples, dictionary learning is performed on
these samples. Being coded by the learned dictionary pair, the
discriminative power of action representations from different
views can be guaranteed in the new feature space, so that the
cross-view action recognition problem can be solved as a traditional supervised learning problem. The proposed approach
achieves state-of-the-arts results on the IXMAS action dataset
using only labeled source view samples, and even outperforms
some methods which utilize correspondence annotations of
action samples across different views. This work leads to a
novel cross-view action recognition setting towards real-world
applications with little information provided.

R EFERENCES
[1] M. Aharon, M. Elad, and A. Bruckstein. K-svd: An algorithm
for designing overcomplete dictionaries for sparse representation. IEEE Transactions on Signal Processing, 54(1):4311
4322, 2006.
[2] J. L. Barron, D. J. Fleet, and S. S. Beauchemin. Performance
of optical flow techniques. International Journal of Computer
Vision, 12(1):4377, 1994.
[3] N. Dalal, B. Triggs, and C. Schmid. Human detection using
oriented histograms of flow and appearance. In European
Conference on Computer Vision (ECCV), 2006.
[4] P. Dollar, V. Rabaud, G. Cottrell, and S. Belongie. Behavior
recognition via sparse spatio-temporal features. In IEEE International Workshop on Visual Surveillance and Performance
Evaluation of Tracking and Surveillance, 2005.
[5] A. A. Efros, A. C. Berg, G. Mori, and J. Malik. Recognizing
action at a distance. In IEEE International Conference on
Computer Vision (ICCV), 2003.
[6] D. Gavrila and L. Davis. 3-d model-based tracking of humans
in action: a multi-view approach. In IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), 1996.
[7] A. Gilbert, J. Illingworth, and R. Bowden. Action recognition
using mined hierarchical compound features. IEEE Transactions
on Pattern Analysis and Machine Intelligence, 33(5):883897,
2011.
[8] I. Junejo, E. Dexter, I. Laptev, and P. Perez. View-independent
action recognition from temporal self-similarities.
IEEE
Transactions on Pattern Analysis and Machine Intelligence,
33(1):172185, 2011.
[9] I. Laptev, M. Marszalek, C. Schmid, and B. Rozenfeld. Learning
realistic human actions from movies. In IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), 2008.
[10] R. Li and T. Zickler. Discriminative virtual views for cross-view
action recognition. In IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), 2012.
[11] Y.-C. Lin, M.-C. Hu, W.-H. Cheng, Y.-H. Hsieh, and H.-M.
Chen. Human action recognition and retrieval using sole depth
information. In ACM International Conference on Multimedia
(ACM MM), 2012.

[12] Z. Lin, Z. Jiang, and L. S. Davis. Recognizing actions by shapemotion prototype trees. In IEEE International Conference on
Computer Vision (ICCV), 2009.
[13] J. Liu, M. Shah, B. Kuipers, and S. Savarese. Cross-view action
recognition via view knowledge transfer. In IEEE Conference
on Computer Vision and Pattern Recognition (CVPR), 2011.
[14] J. Liu, Y. Yang, and M. Shah. Learning semantic visual
vocabularies using diffusion distance. In IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), 2009.
[15] F. Lv and R. Nevatia. Single view human action recognition using key pose matching and viterbi path searching. In
IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), 2007.
[16] J. Mairal, F. Bach, J. Ponce, and G. Sapiro. Online dictionary
learning for sparse coding. In International Conference on
Machine Learning (ICML), 2009.
[17] M. Raptis and S. Soatto. Tracklet descriptors for action
modeling and video analysis. In European Conference on
Computer Vision (ECCV). 2010.
[18] P. Scovanner, S. Ali, and M. Shah. A 3-dimensional sift
descriptor and its application to action recognition. In ACM
International Conference on Multimedia (ACM MM), 2007.
[19] L. Shao, S. Jones, and X. Li. Efficient search and localisation
of human actions in video databases. IEEE Transactions on
Circuits and Systems for Video Technology, 24(3):504512,
2014.
[20] G. Sharma, F. Jurie, C. Schmid, et al. Expanded parts model
for human attribute and action recognition in still images. In
IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), 2013.
[21] Y. Song, L.-P. Morency, and R. Davis. Action recognition by
hierarchical sequence summarization. In IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), 2013.
[22] H. Wang, A. Klaser, C. Schmid, and C. Liu. Action recognition
by dense trajectories. In IEEE Conference on Computer Vision
and Pattern Recognition (CVPR), 2011.
[23] J. Wang, J. Yang, K. Yu, F. Lv, T. Huang, and Y. Gong.
Locality-constrained linear coding for image classification. In
IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), 2010.
[24] Y. Wang and G. Mori. Hidden part models for human action
recognition: Probabilistic versus max margin. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(7):1310
1323, 2011.
[25] D. Weinland, E. Boyer, and R. Ronfard. Action recognition from
arbitrary views using 3d exemplars. In International Conference
on Computer Vision (ICCV), 2007.
[26] S. Xiang, F. Nie, Y. Song, and C. Zhang. Contour graph
based human tracking and action sequence recognition. Pattern
Recognition, 41(12):36533664, 2008.
[27] A. Yao, J. Gall, and L. Van Gool. Coupled action recognition
and pose estimation from multiple views. In IEEE International
journal of computer vision (ICCV), 2012.
[28] Z. Zhang, C. Wang, B. Xiao, W. Zhou, S. Liu, and C. Shi.
Cross-view action recognition via a continuous virtual path. In
IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), 2012.
[29] J. Zheng, Z. Jiang, P. Phillips, and R. Chellappa. Cross-view
action recognition via a transferable dictionary pair. In British
Machine Vision Conference (BMVC), 2012.
[30] F. Zhu and L. Shao. Weakly-supervised cross-domain dictionary learning for visual recognition. International Journal of
Computer Vision, 2014.
[31] F. Zhu, L. Shao, and M. Lin. Multi-view action recognition
using local similarity random forests and sensor fusion. Pattern
Recognition Letters, 34(1):2024, 2013.