Impact of Shifting Time-Window Post-Processing on the Quality of Face Detection Algorithms

We consider binary classification algorithms, which operate on single frames from video sequences. Such a class of algorithms is named OFA (One Frame Analyzed). Two such algorithms for facial detection are compared in terms of their susceptibility to the FSA (Frame Sequence Analysis) method. It introduces a shifting time-window improvement, which includes the temporal context of frames in a post-processing step that improves the classification quality. Error measures are proposed to express the frame-wise accuracy of classifying algorithms, as well as the segmentation of the result sequences which they produce. The two compared algorithms, after applying the FSA improvement, perform better in terms of all the considered measures. The performed experiments have allowed to draw conclusions regarding preferred methods of measuring accuracy of such algorithms and the selection of suitable classification algorithms for being improved. In the end of the work, the resulting future possibilities of further developing the FSA methods are noted.