Encoder settings impact on intra-prediction-based descriptors for video retrieval

Intra-prediction modes (IPMs) of H.264/AVC, as visual feature components, are easily extracted from the compressed domain. However, in spite of their efficiency and effectiveness, they can be sensitive to encoder settings such as encoder-type and encoding-profile. In content-based video retrieval using different encoders between a video and its copy can affect performance. Similarly, applying different encoding-profiles can also have a negative impact. Experiments with different IPM-based descriptors, on a subset of the TRECVID/CCD 2011 collection with content-preserving visual distortions, compare the performance of four H.264/AVC encoding profiles (High10, High, Baseline and Main) and also investigate the impact of three encoders (FFmpeg, MediaCoder and JM). The research shows that ordinal pattern distribution of unified IPMs can improve their tolerance to changes in encoder settings. Furthermore, it is shown that a Correlogram rather than a Histogram of the IPMs is more robust with respect to the impact of encoding-profiles and encoder-types.