Evaluating Spatio-Temporal Parameters in Video Similarity Detection by Global Descriptors

The role of partitioned colour-based global descriptors is well known in video similarity detection tasks for their inexpensive yet effective performance compared to local descriptors. They provide robust and discriminative results in content-preserving visual distortions such as strong re-encoding, pattern insertions and photometric effects. The current research evaluates the effectiveness of three spatio-temporal parameters in video similarity detection tasks. The investigated parameters are specifically colour space, frame partitioning and sampling frame rates. CRIM method (video only) is selected as the base due to its optimum performance in content-preserving visual distortions in the TRECVID/CCD (Content-based Copy Detection) 2011. An amended version of CRIM, based on normalised-average luminance is introduced to compare the results with the baseline. The performance comparison is conducted using a subset of the TRECVID/CCD 2011 dataset, affected by four types of content-preserving visual distortions: T3, T4, T5 and T6. The experimental results shows that the normalised-average luminance descriptors offer more robust and competitive performance. Although they yielded a slightly better performance at the highest sampling frame rate (all frames), compared to the baseline, they offer significantly better performance at the lower sampling frame rate. The experimental evidence also reveals that the core competency of the luminance-based descriptors is significantly improved in terms of mean processing time. This metric is generally known as a shortcoming in video processing algorithms. The effect of the number of partitions is also investigated and it has been shown that increasing the number of partitions can severely lower the efficiency of the method, without yielding a significant increase in the performance.