| Title |
A Study of Object Detection on Vision-transformer-based One-stage Detector for Threat Inspection of Passenger Baggages |
| Authors |
정지욱(Ji-Wook Jeong) ; 송윤선(Yoonseon Song) ; 이수열(Sooyeul Lee) |
| DOI |
https://doi.org/10.5573/ieie.2022.59.6.91 |
| Keywords |
Object detection; Vision transformer; One stage detector; Security inspection |
| Abstract |
In this paper, one-stage detector with a Swin transformer backbone is investigated to alleviate the heavy load from the threat inspection of passenger baggages using the x-ray images. As a benchmark test, SIXray-10 dataset is used. As a baseline one-stage detector, PAA (Probabilistic Anchor Assignment), as a state-of-the-art (SOTA) detector, is considered and several modifications with several training loss functions and attention modules are applied to improve the threat screening probability from the baggages. Experiments reveals that if we combine the FreeAnchor loss function and Dynamic Head module with the baseline model, we can get threat detection accuracy of 64.0 mAP on SIXray-10 dataset. |