Multimedia information retrieval, key-frame extraction is a common need-based problem-specific information retrieval technique in video content. This information retrieval technique reduces the computational power by eliminating a huge set of unwanted frames through the removal of unnecessary content and retrieves key feature frames based on specific video analysis problems. In video analysis classification and captioning of video content using deep neural models, spatio-temporal clues are the most important features. In these two features, modeling of spatial feature is an object understanding by its spatial background, which requires a huge set of frames because objects are occluded by temporal, moving objects. Our instance-based key-frame extraction techniques solve this problem by extracting clear background frames in a video sequence. Here we use a popular PointRend instance segmentation approach to find the moving instances in a video sequence. Effective Instance Prediction Matrix (EIPM) computes predicted instances, which are then evaluated to find the least minimum instance occluded key-frame. We discuss the subdivision techniques of PointRend instance segmentation and results on video datasets. We present results on instance-based key-frame extraction with different datasets and neural models.

Instance Segmentation, Occlusions, Deep Learning, Background Extraction, Key Background