|
Object detection is a common task for defense and intelligence applications in the visible spectrum. Deep learning models achieve state-of-the-art performance for this task. Besides, thermal cameras are more robust against real-world computer vision problems such as illumination changes. Thus, they are commonly used in military applications and security surveillance tasks. Recently, significant improvements have been made by exploiting numerous data for object detection. However, collecting and labeling a vast number of samples is challenging, especially for applications on the thermal spectrum. Since deep networks are sensitive to domain shift, a deep model trained on visible spectrum data may fail to generalize thermal spectrum data. Therefore, developing a method to adapt models to different domains, e.g., visible-to-thermal, is crucial. Feature-level unsupervised domain adaptation methods train a model that maps both domains into a common feature space without requiring image pairs. Such methods reduce the domain gap between two different domains. Besides feature-level adaptation, we propose applying pixel-level transformations on the source domain, e.g., visible spectrum images, to reduce further the domain gap between the visible and thermal spectrum. This results in a significant improvement in the performance of domain adaptive object detection methods. We propose to apply gray-scale conversion, histogram matching, histogram equalization, gamma correction, adaptive gamma correction, and Fourier domain adaptation as pixel-level transformations on visible spectrum images. We conduct experiments on real-world datasets. The evaluations use Cityscapes as the visible spectrum dataset and FLIR ADAS as the thermal spectrum dataset.
|
