| dc.description.abstract | The registration of legal and illegal firearms takes up a substantial amount of
man-hours, in addition to requiring expert knowledge of firearm brands and
subtypes. Therefore, Dutch police force are looking into an alternative way to
classify of firearms which would facilitate the registration process. No
dataset containing brand and subtype annotation was yet available.
Consequently, the current study constructed three firearm datasets containing
brand and subtype annotation, through the means of web scraping and
photo shoots at one of the Dutch police forces depots. Due to data shortage,
visual similarities of the firearms and the wide range of applicable situations
required by the Dutch police force, three approaches were formulated in order
to implement an effective prototype. These approaches were: a baseline approach
that experimented with the effect of data augmentation and image size, a
fine-grained approach that tested the novel combination of two fine-grained
loss functions to increase the attention for fine-grained details, and finally,
an object detection approach that added an object detection pipeline to the
classification algorithm. Subsequent implementation of each approach resulted
in increased performances, achieving a final balanced accuracy score of 80.22\%
on the main benchmark dataset. Taken together, this study revealed that
automating the process of firearm brand and subtype classification is feasible. | |
