IEEE Consumer Electronics Magazine - September/October 2022 - 60

An Incremental Learning of YOLOv3
weighted loss between the distillation loss and
training loss. The l balances the distillation loss
and the normal YOLOv3 loss. The higher value of
l will increase the accuracy of current classes
while a lower value of l will increase the accuracy
of currently learning classes. If l ¼ 0 exhibits that
there is no distillation applied. This presents the
catastrophic forgetting scenario.
Incremental Learning Through Gradient
Descent Memory
In this section, we propose GEM to train
model A on DO for classes CA. The main feature of
this method is episodic memoryMA, which stores
a subset of trained classes CA from model A. In
the incremental stage such that training model B
with new classes CB from data DN, and the stored
memory MA. While training model B, gradients g1
and g2 are calculated. The gradient g2 is updated
when the dot product betweeng1 andg2 is greater
or equal to zero, otherwise, the gradient of the
model B is not updated. The gradients are projected
to the closest gradient in l2 normalization.
EXPERIMENTAL RESULTS
In this section, we extensively evaluated the
proposed incremental learning method on the
PascalVOC 2007 detection benchmark dataset.
Since the focus of this article is on incremental
learning therefore the model is trained only for
two classes: cars and persons. The PascalVOC
2007 contains 713 images of the car class and 804
images of the person class. The model A is trained
on the car class and the model B is trained on the
person as an incremental stage. For the model
accuracy evaluations, we use the mean average
precision (mAP) at the threshold of 0.5 intersection
over the union as the evaluationmetrics.
Implementation Details
We use stochastic gradient descent for training
both the models A and B. The model A and B
are trained for 300 iterations. We used the train
and val sets for training. The test set is used for
testing of the PascalVOC 2007. For training the
model B, one extra output layer is added as we
increased the number of classes. However, the
intermediate layers are kept the same. The model
is trained with different learning rates such as
0.0001, 0.05, and 0.001. However, it is observed
that 0.001 is the optimal learning rate for training
60
Table 1. Hyperparameters for Training Model A and
Model B.
Parameters
Learning rate
Momentum
Weight decay
Iteration
Batch size
Image size
Values
0.001
0.937
0.0005
300
2
416
themodel. Themodel training parameters are presented
in Table 1.
Incremental Learning Through Knowledge
Distillation
To validate the performance of the proposed
incremental learning method, we trained the
model A on the car class only as normal training.
The model B is incrementally trained on the person
class while only utilizing the labels for the
person class. The annotations of the car class
are ignored.
We evaluated the performance of the proposed
incremental learning by utilizing auxiliary
data and without using auxiliary data. The new
data DN contain the labeled person images and
raw images. These raw images are utilized as auxiliary
data that contain the unlabeled previously
learned objects. The baseline approach of adding
new classes is to add more output layers in the
YOLOv3 model. The model B inherits the pretrained
weights of the model A trained on the car
class. When training on the model B, the model A
weights are utilized to predict the previously
learned class in new training dataDN. This prediction
of the previous class is updated in the output
layer of the model B. The experiments are performed
with auxiliary data and without auxiliary
data. The model B is trained with data that contains
unlabeled previous classes known as auxiliary
data. If there is no auxiliary data while
training the model B, the average precision drops
down to 50% as shown in Table 2. Table 2
presents the accuracy of normal training, the
model A, catastrophic forgetting, the model B
with and without using auxiliary data. The graphical
representation of all trainings presented in
Table 2 is shown in Figure 3. The YOLOv3 model
IEEE Consumer Electronics Magazine
IEEE Consumer Electronics Magazine - September/October 2022

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - September/October 2022
