Performance Evaluation of Text Detection and Tracking in Video

Abstract

Text detection and tracking is an important step in a video content analysis system as it brings important semantic clues which is a vital supplemental source of index information. While there has been a significant amount of research done on video text detection and tracking, there are very few works on performance evaluation of such systems. Evaluations of this nature have not been attempted because of the extensive effort required to establish a reliable ground truth even for a moderate video dataset. However, such ventures are gaining importance now. In this paper, we propose a generic method for evaluation of object detection and tracking systems in video domains where ground truth objects can be bounded by simple geometric shapes (polygons, ellipses). Two comprehensive measures, one each for detection and tracking, are proposed and substantiated to capture different aspects of the task in a single score. We choose text detection and tracking tasks to show the effectiveness of our evaluation framework. Results are presented from evaluations of existing algorithms using real world data and the metrics are shown to be effective in measuring the total accuracy of these detection and tracking algorithms.

Multi Cue Performance Evaluation Metrics for Tracking in Video Sequences

“…However, the approach had a very restricted application range with the focus being on detecting and tracking text in video which involves objects of consistent size and shape across a sequence. We adopt and improve this approach [15] by also concentrating on the details inside a bounding box, i.e. all the details inside a bounding box are marked ground truth and is the region of interest (ROI).…”

“…Such methods could give good results on single sequences. A more generic approach in the sense that it could combine multiple features for evaluation has been described [15]. The work involved the use of two comprehensive measures for text object detection and tracking systems for which the ground truth objects are bounded by simple geometric shapes.…”

Ανίχνευση Κειμένου Σε Εικόνες Και Βίντεο

“…Σύμφωνα με άλλες μεθόδους αποτίμησης, το στοιχείο προς ανάκτηση δεν είναι το εικονοστοιχείο αλλά το κουτί κειμένου. Στις σχετικές εργασίες[45] και[46] οι Kasturi και Manohar και συν. προτείνουν ως συνολική μετρική…”