Semantic segmentation and description for video transcoding

Abstract

We present an automatic content-based video transcoding algorithm which is based on how humans perceive visual information. The transcoder support multiple video objects and their description. First the video is decomposed into meaningful objects through semantic segmentation. Then the transcoder adapts its behaviour to code relevant (foreground) and non relevant objects differently. Both objectbased and frame-based encoders are combined with semantic segmentation. Experimental results show that the use of semantics and description prior to transcoding reduces the bandwidth requirements and makes it possible to adapt the video representation to limited network and terminal device capabilities still retaining the essential information.

Semantic video analysis for adaptive content delivery and automatic description

“…Content-blind transcoding strategies include spatial resolution reduction, temporal resolution reduction, and bit-rate reduction [10]. Recent transcoding techniques make use of semantics to minimize the degradation of important image regions [11], [12]. In [13], optimal quantization parameters and frame skip are determined for each video object individually.…”

A Framework for Ontology Enriched Semantic Annotation of CCTV Video

“…However, the large numbers of installed CCTV cameras that operate continuously produce enormous video datasets, in which users need to search to access video shots that are semantically interesting. Therefore, video footages need semantic transcoding [1], i.e. indexing of video shots with the appropriate semantic labels.…”

Computer vision techniques for PDA accessibility of in-house video surveillance

“…Most of non-trivial background models use statistical function over the history set of frames: Mode [18], Median [10], Mixture of Gaussians [23,20], PCA reduction [15], and so on, have been proposed. Moreover, the background model is often selectively updated according with the knowledge extracted from previous frames: for instance, pixels or objects that have been detected in motion in previous frames are not used to update the background [7,2].…”

“…1 says that the new background B t+∆t (p) can be in some pixels equal to the previous B t (p) if we have knowledge of that pixel as a foreground one, equal to the current frame I t+∆t (p) if some events occur, or computed statistically as B t s (p). Bs is 2 Examples are objects in the background like a door that starts its motion and leaving a "ghost" blob in its initial position. Ghosts and MVOs are discriminated with a validation step that uses the average difference of the object's pixels with previous frame.…”