Instrumentation & Measurement Magazine 24-3 - 80

Fig. 2. Flow diagram of multi-face tracking system.

classification methods. In our study, we experimented with
various classic object detection meta-architecture, and we will
consider the following two methods further:
◗◗ SSD [11] uses low-level feature to localize the bounding
boxes and extracts high-level feature to classify objects
without extra work to generate region of interests.
◗◗ R-FCN [12] shares almost all computation on the entire
image. It achieves good accuracy in Huang et al. [10], but
it still takes considerable time compared to SSD.
Numerous common feature extractors can be utilized in object detectors, and we use the following classic extractors:
◗◗ Inception V2 [25] decomposes large kernels to small
kernels and reduces computational cost by replacing
large kernels multiplication into small kernels multiplication and addition.
◗◗ Residual learning and residual network (ResNet) [26]
uses identity mapping through shortcut connections
between ReLU layers to avoid gradients exploding or
vanishing.
◗◗ MobileNet [27] can reduce the CNN computational cost
based on separable depth-wise convolutional filters
and point-wise convolutional filters in the width multipliers, and resolution multipliers reduce the image
sizes.

Deep Association Metrics and Tracking Matching
Cascade
The deep association metrics for multi-object tracking (Deep
SORT) [3] combines both physical motion and feature appearance information to overcome the occlusion problem,
and it decreases the rate of IDS, i.e., the number of times
the label for the same face changes. The physical motion
information is compared with the Mahalanobis distance.
The calculation of the distance compares the predicted
Kalman states based on previous video frames and face
80

measurements in the current frame. For tracking through occlusions over a long period of time, a second metric based on
feature appearance is used. This second metric measures the
smallest feature distance between an active face track and
a face detection in the current frame in feature appearance
space. Wojke et al. [3] combine these metrics in the case of pedestrians with a matching cascade. In this cascade, the two
metrics are used separately as a gate to find possible matches
and combined in a weighted cost function to calculate the
matching cost between an active face track and a face detection in the current frame.

Classifier Loss Functions
Different loss functions have been considered in face detection
and face ID matching. Wojke et al. [4] have proposed a Cosine
Softmax Classifier loss for pedestrians [4]. We will compare
the same loss functions but for faces with the Angular Softmax Loss [5] which has been specifically proposed for faces.
We summarize the Cosine and the Angular Softmax Classifier
loss functions below.
Cosine Softmax Classifier: The Cosine Softmax Classifier
can be derived from the Multi-Class Softmax Classifier by
removing the bias terms for all classes. Next, a free scaling parameter κ is added to be the marginal factor for each class.
Additionally, the weights are normalized to unit length, i.e.,
 k  wk / w k ,  x  1,..., C [4]. Therefore, the Cosine Softmax
w
2
Classifier is:

∣r 

p  y k



 Tk r
exp   w







 Tn r
exp   w

n1



. (1)

Lower κ creates a less discriminative margin, and higher κ
places larger penalty on misclassified samples to give higher
discrimination.

IEEE Instrumentation & Measurement Magazine

May 2021

Instrumentation & Measurement Magazine 24-3

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 24-3