| Summary: | 碩士 === 國立高雄科技大學 === 電腦與通訊工程系 === 107 === Ground truth annotation is crucial in evaluating the performance of pedestrian detection algorithms. Nevertheless, manually annotation is a tedious and time-consuming task. In this work, we propose a pedestrian labeling system based on on-line boosting for transfer learning with multiple sources to assist users to efficiently annotate the ground truth of dataset consisting of a sequence of images. The proposed system mainly consists of four steps. Firstly, the system generates proposal annotations by using pedestrian pre-detector and prediction based on temporal information. Next, pre-classifier confirms the proposal annotations so as to effectively preserve the possible pedestrian candidates which is called initial annotations. Then, the correct annotations are obtained by the intervention of users. Finally, using the obtained correct annotations, online training mechanism is applied to update the pre-classifier and transfer the domain knowledge from multiple sources to the target datasets. In the experiment, we decompose annotation actions into two primitive actions including click and drag. The metrics for evaluating the annotation effort are click number and drag distances. The experimental results demonstrate that the proposed system effectively reduces the annotation cost.
|
|---|
