6 Experimentos y análisis de resultados En este capítulo se abordan y estructuran todos los experimentos
6. Preguntas temporales que contienen entidades y varias fechas. Por ejemplo, “Nombre una película en la que haya participado Kirk Douglas entre 1946 y
6.4.5 Evaluación de BRUJA atendiendo al tipo de pregunta
Mismatch problem in predictive multiple description video encoders can be avoided completely or partially by including mismatch coding using the different meth- ods discussed in Section 2.6.1.1. A very simple example to partially control the mismatch is to use the intra coded frames in each description at regular intervals to reset the mismatch errors. Another approach to control mismatch is to use three prediction loops as shown in the diagram of generic multiple description predictive encoder.
In [95], two different algorithms are proposed that solves the mismatch problem partially and completely by using three prediction loops at the encoder. In both the algorithms, motion vector and header information is duplicated in each de- scription. Intra and motion compensated prediction error frames are coded into multiple descriptions by using PCT. In algorithm 1, mismatch is avoided com- pletely by embedding the side prediction error in each description. Similarly, in second algorithm mismatch is controlled partially by embedding partial amount of side loop prediction error. Therefore, results into reduction in redundancy
between the descriptions.
2.6.2
Scalable Multiple Description Video Coding
The MDVC schemes discussed in the previous section are based on non-scalable predictive video coding framework. The advantage of those schemes is that one can use the existing standardized coding blocks in MDC framework, which are designed for high coding gain. However, high coding gain is not always the only requirement especially in case of heterogeneous bandwidth requirements. Scal- able coding addresses the problem of heterogeneity but is prone to errors. On the other hand, MDC is a source coding scheme, which is robust against transmis- sion losses. Therefore, different schemes that combine the scalable and multiple description coding are proposed to provide both the scalability and reliability. Scalable multiple description video coding (SMDVC) schemes can be categorized into two categories.
2.6.2.1 3D Transform based Scalable Multiple Description Video Cod-
ing
In this section, open loop architecture of SMDVC based on 3D transform is dis- cussed. The open loop architecture uses the MCTF and multi resolution spatial decomposition that provides the scalability and gives better performance than the close loop scalable video coding [103]. A general framework of the 3D transform based MDVC is shown in Figure 2.9. Firstly, the temporal correlation among different frames is removed by applying MCTF. The MCTF block converts the input video into low and high frequency frames and a set of motion vectors. The low and high frequency frames are then be decomposed by 2D spatial transform to remove the spatial redundancy and hence complete the 3D saptio-temporal decomposition. Different methods are available in the literature on generating multiple descriptions from the spatio-temporal decomposed coefficients.
In [10], Bajic and Woods proposed a MDC method for images and video based on data partitioning. Multiple descriptions are generated for images by partitioning
MCTF 2D-DWT Multiple Description Encoder Input
Video
Motion Vectors
Figure 2.9: 3D transform based multiple description video coding framework.
the wavelet transformed coefficients in such a way that each description contains the coefficients that are maximally separated from each other. By using data partitioning no redundancy is added in each description and errors can be con- cealed from the received descriptions. In case of video, the 3D spatio-temporal coefficients and the motion vectors are partitioned into different descriptions in a similar way as in images. In [11], multiple descriptions are generated from the spatio-temporal coefficients by repeating the low frequency frames and motion vector information in each description and high frequency frames are divided between the two descriptions. The redundancy is controlled by the amount of duplicated information in each description. In case of single description decoding, the missing frames are estimated from the motion vector information, therefore the quality of single description decoding is low for sequences having high motion.
Another 3D wavelet transform based scheme that generates flexible number of descriptions is proposed in [104]. The main concept of this scheme is to encode each spatio-temporal decomposed code block to any given number of descriptions. Each code block is encoded at high data rate for one description and at low data rate for all other descriptions. Different code blocks are then mixed from high data rate and low data rate encoding to generate a description. A rate scalable coding scheme (EBCOT) is adopted to encode each code block. If all the descrip- tions are available at the decoder, the code blocks encoded at high data rate from each description are selected for decoding. On single description decoding, still an acceptable but low decoding quality is possible from the low data rate code blocks. The redundancy is controlled by the rate-distortion selection of each code block. In [12], embedded MDSQ proposed in [54, 82, 83] is used to generate scal- able descriptions from the spatio-temporal decomposed coefficients. The channel aware rate allocation algorithm is also adopted in this scheme to further improve the reliability of the video delivery under packet and bursty losses. Inspired from the results of MDSQ-SR in SMDC for images, a 3D transform based framework
for scalable multiple description video coding is proposed in Chapter 6 that uses the concept of MDSQ-SR. The scalable descriptions of the video generated from several MDSQ-SR are capable of joint decoding at different quality, resolution and frame rate in balanced and unbalanced manner. Two schemes that reduces the texture and motion vector redundancies are also proposed in Chapter 6.