Joanneum Research at TRECVID 2005 – Camera Motion Detection

Low-level feature extraction (camera motion) Ground truth annotation Manual camera motion annotation has been performed by three groups using a tool provided by Joanneum Research. Some types of content made it difficult or impossible for human annotators to describe the camera motion. The comparison of annotations of the same content done by two groups shows significant differences for some features. We discuss the questions that arise from the results of the manual annotation. Submitted runs Our approach is based on feature tracking, clustering the feature trajectories and selecting the cluster representing dominant motion. The parameters for the decision about the presence of certain types of camera motion from the estimated motion parameter sequence differ between the runs. JRS1: same parameters for pan and tilt, stricter parameters for zoom JRS2: slightly reduced parameters for tilt, reduced parameters for pan and zoom Modification of pan and tilt parameters increased recall by slight loss of precision, the tradeoff between precision and recall was higher for zoom. The decision about the presence of a certain type of camera motion from the estimated motion parameter sequence is most critical point. Cross-evaluation between parameters trained on the development or test data set on the respective other set should be done. 1 Common Annotation of Camera Motion Ground Truth