The following work proposes a benchmark of performances of state of art AI algorithms for the weapons detection. Particularly, it is aimed to test three CNN based models on the task of detecting specific types of weapons. In order to accomplish this goal, four datasets are employed. Additionally, due to the lack of rich amounts of well-structured datasets in these field of research, new labeled data are produced as a new resource to test specific hypotheses about their impact on the performances of the models: different transfer-learning approaches are studied to understand how specific types of data could increase the generalization of the trained algorithms, with a peculiar attention to realistic scenarios. The whole work is designed with the intention to select the state-of-art algorithms that truly could be employed for realistic applications
Comparing Artificial Intelligence Algorithms in Computer Vision: The Weapon Detection Benchmark
Dentamaro V.
;Giglio P.
;Impedovo D.;Pirlo G.
2022-01-01
Abstract
The following work proposes a benchmark of performances of state of art AI algorithms for the weapons detection. Particularly, it is aimed to test three CNN based models on the task of detecting specific types of weapons. In order to accomplish this goal, four datasets are employed. Additionally, due to the lack of rich amounts of well-structured datasets in these field of research, new labeled data are produced as a new resource to test specific hypotheses about their impact on the performances of the models: different transfer-learning approaches are studied to understand how specific types of data could increase the generalization of the trained algorithms, with a peculiar attention to realistic scenarios. The whole work is designed with the intention to select the state-of-art algorithms that truly could be employed for realistic applicationsI documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
