Feature-Cut: Video object segmentation through local feature correspondences

Accurately segmenting objects in video is a difficult and time consuming process in modern post-production houses. Automatic systems may work for a small number of frames, but will typically fail over longer video shots. This work proposes a semi-automatic, feature-based system to perform object segmentation over longer sequences. The user manually extracts masks from representative instances of the object, which are then propagated to the remaining unsegmented frames and used to bootstrap the automatic segmentation for these frames. The presented work dramatically reduces the manual workload required to segment a video sequence, allowing longer and more accurate object mattes.