Infrared Pedestrian Dataset Training using Swin Transformer model
Yeonha Shin, Sungho Kim · 2022 22nd International Conference on Control, Automation and Systems (ICCAS) · 2022
Recently, studies to replace CNN-based models with Transformer models are being actively conducted. In this paper, the Swin Transformer model, which has recently been attracting attention for its excellent performance, was trained using infrared images and its performance was examined. A Swin Transformer designed for RGB images was tuned for infrared image training to train a separate infrared pedestrian dataset and a simple experiment was performed to further improve the infrared image training performance. In addition, in order to properly tune the Swin Transformer Backbone model to the infrared image data and to improve the training performance, we trained separately configured RGB and infrared datasets and analyzed the results. As a result, it was concluded that the Swin Transformer would be suitable for infrared data training if it was slightly tuned for infrared data and lightweight to avoid overfitting. Based on this experiment, the author plans to create a model suitable for infrared datasets in the future and apply it to various practical applications.