IEEE Consumer Electronics Magazine - September/October 2022 - 62

An Incremental Learning of YOLOv3
Figure 4. Object detection of test images on the proposed method. (a) Model A. (b) Model B.
method. However, the detection time of FRCNN5 is
higher than the proposed method. The detection
time of the proposed method is lower among all.
Few images from the testing images are predicted
with model A and model B as shown in Figure 4. In
Figure 4(a), the model only predicted car class. In
Figure 4(b), both classes are predicted despite the
fact that car labels are not used in training. This
indicates that the proposed method has retained
the previous knowledge of model A.
Incremental Learning Through Gradient
Episodic Memory
The GEM is originally implemented on image
classification tasks.7 We implemented the GEM
on the YOLOv3 object detection model for comparison.
Table 3 illustrates the accuracy of two
classes of the PascalVOC 2007 dataset. The car
class is trained as normal training while the person
class is trained as an incremental stage. The
GEM performs better to prevent the previous
class from forgetting however it lags in current
class accuracy in comparison to DMC12 and
FRCNN.11 The mAP of GEM is 7.7% higher than
DMC but 15.45% and 16.15% lower than FRCNN11
and the proposed method, respectively.
CONCLUSION
In this article, we introduced the problem of
incremental learning for object detection without
62
utilizing the previous training data. A knowledge
distillation loss is proposed for the prediction of
the previous task combined with YOLOv3 losses
to optimize the performance of current classes.
The proposed method does not utilize the images
regarding previously learned classes making it
appropriate for strict privacy and storage
requirements. Additionally, the previous model
weights are inherited to update the new one, thus
minimizing the memory requirement. The proposed
method has largely alleviated the problem
of catastrophic forgetting, which proved to be
critical in the case of object detection. Extensive
results on the pascalVOC 2007 dataset show that
the proposed method significantly increased the
mAP up to 74.05% for two classes. Additionally,
the proposed method achieved the lowest runtime
among other incremental learning methods.
&REFERENCES
1. V. K. Kukkala, J. Tunnell, S. Pasricha, and T. Bradley,
" Advanced driver-assistance systems: A path toward
autonomous vehicles, " IEEEConsum. Electron. Mag.,
vol. 7, no. 5, pp. 18-25, Sep. 2018, doi: 10.1109/
MCE.2018.2828440.
2. S. F. Cotter, " MobiExpressNet: A deep learning network
for face expression recognition on smart phones, " in
Proc. IEEEInt. Conf. Consum. Electron., Jan. 2020,
pp. 1-4, doi: 10.1109/ICCE46568.2020.9042973.
IEEE Consumer Electronics Magazine

http://dx.doi.org/10.1109/MCE.2018.2828440 http://dx.doi.org/10.1109/MCE.2018.2828440 http://dx.doi.org/10.1109/ICCE46568.2020.9042973

IEEE Consumer Electronics Magazine - September/October 2022

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