A New Dataset and Evaluation for Infrared Action Recognition

Action recognition (AR) is one of the most important tasks in computer vision and there are a large number of related research works along this line. While most of these works are investigated on AR datasets collected from the visible spectrum, the AR problem on infrared scenarios still has not attracted much attention, and there is even few public infrared datasets available for supporting this research. This study aims to emphasize the importance of the infrared AR problem in real applications and arouse researchers’ attention on this task. Specifically, we construct a new infrared action dataset and evaluate the state-of-the-art AR pipeline, including widely-used low-level local descriptors, coding methods and fusion strategies, on it. Through these evaluations, we find some interesting results. E.g., dense trajectory feature can achieve the best performance while the appearance features, e.g., HOG, has relatively poorer performance; the coding method of vector of locally aggregated descriptors is evidently better than that of the widely-used fisher vector; the late fusion facilitates a better performance than early fusion. Furthermore, the best performance achieved on our dataset is 70%, leaving a relative large space for promoting new methods on this infrared AR task.