SEGURIDAD NACIONAL COMO CAUSA DE EXPROPIACIÓN

Stereoscopic content contains two images or streams, one for each eye, hence, doubling the amount of data compared to a traditional 2D representation. As current infrastructure (e.g. bandwidth, medium sizes) was primarily aimed at 2D images and videos, stereoscopic data should be compressed in order to facilitate its processing and transmission. This would also help the transition from 2D to

stereo technology and speed up its adoption by the community. While both the left and right views can be compressed separately, a good coding algorithm should exploit the fact that both views are highly correlated with small differences which result from camera separation. This section describes the more popular formats for storing stereoscopic data. Gorley (2012) offers a similar overview in his thesis. Digital stereoscopic imaging is relatively novel so file formats are still mostly unestablished and new stereo standards are still emerging. Current stereoscopic file formats are mostly based on existing image and video formats which get extended to store two or more views. As extensions, these formats inherit prop- erties of the original formats, including the type of compression (e.g. lossy or lossless), coding methods (e.g. discrete cosine transform) and meta information (e.g. colour statistics). Another advantage of extending existing 2D file formats is backwards compatibility. Stereo images usually contain tags or flags in the file description which indicate to the application that data about to follow represents a second view. Legacy applications are unable to recognise these tags, which are ignored and only a single view gets displayed. However, while enabling compati- bility and simple implementations, these file formats do not implicitly take into account similarity between stereo pairs.

2.5.1

Animated Graphics Interchange Format (GIF)

GIF (CompuServe Incorporated, 1990) is a bitmap image format which stores images using up to 8 bits per pixel (bpp). These values identify up to 256 colours

which are stored in a separate colour map. GIF uses the lossless LZW coding

algorithm to compress data. A small size is one of the main advantages of GIF, which made it popular with internet users. Also, it can store number of frames thereby enabling short animations (Animated GIF).

The animation feature is especially interesting for stereoscopy, where left and right views can be stored as consecutive frames which alternate when viewed. The technique, informally termed Wobble GIF, results in motion parallax and produces the sensation of depth. Standard web browsers or image viewers open animated GIFs thereby enabling delivery of stereo to a wide audience without the need for special hardware or software. However, binocular vision is not achieved so depth cannot be fully appreciated and animation can distract from the actual content of the image. The number of colours is limited and GIF is unable to reproduce more complex scenes with high quality

2.5.2

Stereoscopic Portable Network Graphics Format (PNS)

Portable Network Graphics (PNG) format was developed to improve upon GIF and circumvent its patent license. It supports 24 bpp (or 32 bpp when using an alpha channel) allowing for a much larger colour palette and usage of RGB colour space (Roelofs, 1999). PNG uses filtering and theDEFLATE coding algorithm to losslessly compress images. The format is separated into a header section which is followed by data segments named chunks. Chunks can be: critical - it needs to be read by any application opening a PNG file; and ancillary - it can be ignored if the application cannot read it. This enables backward compatibility. Chunks indicated by the letters “sTER” identify a stereo images which are saved in a side-by-side manner. One bit is used to indicate whether the left image comes first. When opened by a supported viewer, the image can be displayed in an appropriate manner, but when opened in a traditional viewer, both images will be displayed side by side. The same behavior, without internal file formatting, can be achieved by saving a PNG file, consisting of a side-by-side stereo pair, using the PNS file extension.

2.5.3

Stereoscopic JPEG (JPS)

Joint Photographic Experts Group (JPEG) format is the most commonly used image format. It uses a lossy method for storing images (lossless is also supported but not used as frequently), achieving higher compression ratios compared to a lossless method with a potential loss of quality (Ghanbari, 2010). Encoding can be implemented in different ways most of which perform the following opera- tions: the image is converted to Y0CBCR colour space and chroma components (CBCR) are downsampled by 2; the image is separated into 8×8 pixel blocks which are transformed using discrete cosine transform. Frequency amplitudes are then quantised based on frequency and finally lossless Huffman coding is used to fur- ther compress data. Similar to PNG, JPEG is separated into segments identified by markers. Stereoscopic JPEG images have the JPS file extension and use an

APP3 (application specific) marker to indicate the format used to save the stereo pair: interleaved, side-by-side, over-under and anaglyph (Siragusa et al., 1997). Viewing behaviour is also similar to PNS, so supported applications will show the image appropriately while traditional viewers will show both images the way they were stored.

2.5.4

Multiview Video Format (MVC)

Advanced Video Coding (AVC), also referred to as MPEG-4 Part 10 or H.264, is a video format which is currently the most popular format for storing videos. It is used by Blu-ray Discs, YouTube, Adobe Flash Player and HDTV broadcasters. AVC consists of a family of standards, which provide versatile functionalities, only a subset of which may be relevant for a specific use (Ghanbari, 2010; Richard- son, 2010). It supports both lossless and lossy coding using techniques includ- ing: wavelets, discrete cosine transform, entropy coding and deblocking filters. Broadly, AVC can be described as a block-oriented motion-compensation coder. Block-oriented refers to the idea of separating the image into blocks which are compressed separately (similar to JPEG). Motion-compensation is a technique for predicting a frame from previous or subsequent frames, based the fact that consecutive frames in video are alike.

As the same fact holds for stereoscopic and multiview images, AVC was extended to include Multiview Video Coding - a specification which also uses motion-compensation to predict the second image of stereo pair. The process is illustrated in Figure 2.11. Frames from two stereo streams are coded using MVC specification. The frame used as a reference is termed an I-frame (intra-coded picture) and is coded independently from any other image. In general,P-frames (predicted picture) store differences from the previous frame, whileB-frames (bi- predictive picture) store differences from previous and following frames. In the stereo case, the other view is also included in the prediction. I-frames are the largest in size as motion compensation is avoided, while B-frames are smallest.

Figure 2.11: Multiview video coding relies on the motion compensation, where some frames are predicted from the previous and the following frames, and from the other view.

However, B-frames imply that sections of video need to be loaded before they are played to enable reconstruction.

The MVC approach is effective in the case of multiview and improves with the number of additional streams compared to coding each stream individually using MPEG-4. However, in the stereo case gains are rather limited, because temporal prediction performs well already and additional inter-view prediction does not significantly affect coding efficiency. A major improvement is achieved in case of I-frames of the second view which get coded as P-frames, but they typically occur every 0.5 - 1 s.