Signal Processing - November 2017 - 82

11100100110
00010011000

10011001011
01001100101

Quantization Loss

learned compact binary codes to exploit nonlinear relationship of
samples, and Figure 6 shows the key idea of their method.

Balanced Bits

Person reidentification
Yi et al. [12] proposed a DML method for person reidentification,
which learned a similarity metric from image pixels directly with a
Siamese networks. Hu et al. [23] proposed a DTML method for cross
data set person reidentification, where the discriminative information exploited from the source domain was transferred to the target
domain with limited labeled samples. Shi et al. [31] proposed a deep
embedding metric method for person reidentification, which used a
moderate positive sample mining method for robust CNN training,
and improved the leaning procedure with a metric weight constraint.

Independent Bits

Binarization

Visual tracking

Gallery Images

Hu et al. [10] employed DML for visual object tracking. Their
DML tracker adopts the marginal fisher analysis criterion to
characterize the separability of the positive samples and negative
samples by maximizing the variance of interclass negative sample pairs. They first learned a multilayer nonlinear feed-forward
neural network to map both the sampled templates and particles
into a discriminative feature space to minimize the intraclass
variations of positive sample pairs and maximize the interclass
variations of negative sample pairs at the top layer of the network. Then, they selected the candidate which is most similar to
the template in the learned deep network and considered it as the
target in the current predicted frame. Experimental results demonstrated that their DML tracker achieved very competitive performance on a challenging benchmark data set. Figure 7 shows
the main procedure of their proposed DML tracker.

FIGURE 6. The basic idea of deep hashing for large scale visual search
[35], which employed a feed-forward neural network to map each gallery
image into a compact Hamming feature space with three criteria.

obtain a flexible nonlinear similarity metric from images, which
have multimodal feature representations via an efficient and scalable online learning method. Their proposed technique achieved
encouraging results on several multi-modal images retrieval tasks.
Wang et al. [9] proposed a ranking-based deep-learning method
for fine-grained image search, which employed a triplet-based
hinge loss ranking function and a multiscale neural network. Their
method outperformed existing hand-crafted features and deep
models in numerous experiments. Liong et al. [35] proposed a
deep hashing approach for large scale visual search. Their method

Cross-modal matching
Liong et al. [13] proposed a deep coupled metric learning (DCML)
method for cross-modal matching. Unlike existing cross-modal

y(2)
w(2), b(2)

+
y(1)

DML
-

w(1), b(1)
y

Frame (t )

Positive and Negative Samples

Distance
Metric

Network Architecture

Feature
Representation

Tracking Result

Frame (t + 1)

FIGURE 7. The key procedure of the DML tracker [10]. This tracker sampled some positive and negative samples to construct a training set and learn a
deep metric with these samples at the t th frame, For the coming (t + 1)th frame, their tracker computed the similarity between each candidate and the
template and find the candidate which has the maximal similarity score as the foreground the (t + 1)th frame.
82

IEEE SIGNAL PROCESSING MAGAZINE

November 2017

Table of Contents for the Digital Edition of Signal Processing - November 2017