A Performance Study on Object Detection of Image and Video

Detecting objects is one of the most important problems in computer vision, including face detection, pedestrian detection and vehicle license detection. This paper proposes a new benchmark framework to evaluate different kinds of detecting problems. A new reasonable assessment metric for evaluating output with labeling information is described. We also give an interactive pre-labeling method to reduce the workload of video manual labeling. It can help us build the ground truth labeling datasets. We present lots of generally used evaluation arguments in labeling problem. Finally, by analyzing our improving facing detection algorithm in several datasets, including public datasets and our private datasets, we help identify future research directions for the area.