Content-based group-of-picture size control in distributed video coding

• Controlling group of picture size in DVC is difficult but important for performance.
• This work proposes how to compute the best sequences of GOP sizes.
• An algorithm is proposed to decide, using SVMs, the best GOP sizes in DVC.
• The achieved rate distortion performance is close to the ideal performance.

Controlling the group of picture (GOP) size in distributed video coding (DVC) is a difficult but important task since it has a direct impact on the coding performance. This paper presents a framework to adaptively control the size of GOPs in a Wyner–Ziv encoder by means of encoder-side decisions based on support vector machines (SVM) that uses simple features extracted from the original video content. To train the SVM, firstly this work proposes how to compute the sequence of GOP sizes with the best rate-distortion performance given the set of GOP sizes that can be used during the encoding process. Then, an algorithm based on the previously trained SVMs is presented to control the actual GOP size each time a new decision can be taken at the encoder. Results show that the proposed algorithm can achieve a rate distortion performance close to the ideal one. Moreover, comparisons with a reference adaptive GOP size selection algorithm in the literature shows gains up to 2 dB PSNR in the best conditions.