Quantitative performance evaluation of video analytics algorithms requires reference image sequences along with ground truth information. Such video datasets are essential to compare newly developed methods with state-of-the-art solutions. With the explosion of video-based applications (security, traffic, etc.) and sensor types (mono, color, near-IR, thermal, etc.), an impressive diversity of benchmark datasets are required. Recently, the combination of thermal and color sensors for outdoor video analytics applications turns out to be a prolific solution. However, very few simultaneous color and thermal videos are publicly available for development and benchmarking.
Over the years, INO has developed a strong expertise in using multiple sensor types for video analytics applications in uncontrolled environment. A huge quantity of videos have been grabbed and stored, especially with the VIRxCam platform that has been permanently installed outdoor to grab image sequences at all time of day, and in all weather conditions. The most interesting characteristic of this hybrid LWIR / color sensor is the accuracy of its pixel-based thermal / color image registration. By sharing a few of these videos, with their corresponding ground-truth annotations, we allow development and benchmarking of computer vision algorithms for outdoor video analytics applications.