Towards Practical Evaluation of Pedestrian Detectors
Despite recent significant advancement in the area of pedestrian detection in images, little effort has been devoted to algorithm evaluation for practical purposes. Typically, detectors are evaluated only on color images. It is not clear how the performance would be affected if other modalitites are used e.g. thermal or near infrared. Also, detectors are evaluated on cropped images that have the same size as training images. However, in practice, detectors are applied to large images with multiple pedestrian in different locations and sizes. To apply a single size pedestrian detector, the input image is, typically, scanned several times with different window sizes. It is not clear how the detection performance would be affected by such multiple-size scanning techniques. Moreover, to implement such a technique, one is faced with a multitude of design choices, each of which potentially affects the performance of the detector. The contribution of this paper is to assess and reason about the differences in detection performance of two state of the art detectors across changes in modality (visible or near infrared), evalution method (on cropped or whole images), the effect of different design choices (resizing features or images, smoothing or not).
