Informational Retrieval Assisted Object Segmentation in Video

Accurate object segmentation in video is difficult. The dynamic nature of the medium causes drifts in the feature spaces traditionally used in segmentation of objects in still images. For example, colour distributions, shape models and motion tracks of objects typically vary and /or deteriorate over time, resulting in the need to explicitly correct the object by hand in every few frames. The presented work exploits recent feature-based object detection work from information retrieval (IR) literature to propagate information from frames that are far apart to greatly reduce the amount of time required to manually correct segmentations.