Extend Thermal Infrared Face (ETIF)

In this study, we extend the existing thermal infrared face dataset (TIF) by improving the diversity and quality of data. More specifically, the new data contains more acquisition periods with significant differences among ambient temperature periods and a slow change in ambient temperature within each period. At the same time, we provide the corresponding visible images of the infrared images to assist in face detection and face depth estimation. For noise reduction, we calculated the average facial temperature of the short-term population and determined upper and lower limits. Any data that deviates from this range will be removed. In this way, the noise introduced by false face detection during data collection is reduced. For De-redundancy, we all know that saving all detected infrared and visible face images during data acquisition results in processing inefficiencies. Therefore, we sampled 5000 images at equal intervals from each period to reduce the data volume. Experiments have shown that this sampling scheme does not affect the training and testing of our model.