ComplexFace: A Public Visible-Thermal Face Dataset with Real-Life Complexity

Jiajin He, Chengxi Dong, Yunqi Cai, Dong Wang

Abstract: Thermal face recognition (TFR) can address the limitations of traditional visible-light face recognition under complex lighting conditions. However, public thermal datasets are scarce, and the recording conditions of the few available datasets are relatively simple. In this paper, we present a new public thermal face dataset ComplexFace. The key feature of the new dataset is that it involves a wide range of real-life complexity, including variations in scenarios, sessions and devices. Comprehensive experiments were conducted with the new dataset to: (1) discover the most influential factors that impact TFR performance; and (2) discover the actual performance of existing TFR models when real-life complexity is involved. Our experiments demonstrated that variations in scenarios, sessions and devices all pose a significant impact on TFR, and the TFR performance of existing models is much worse than reported in the literature under real-life conditions.