Automatic extraction of salient objects in 3D stereoscopic videos

For 3D stereoscopic videos the depth perception represents an important factor that affects the human visual attention much more than any motion or texture contrast existent in a traditional 2D videos. In this context, the present paper addressed the issue of stereoscopic visual attention models designed to detect salient objects in 3D videos. We propose representing the image sequence as a 2D video stream and its associated depth maps. The technique starts by combining a spatiotemporal attention model with a disparity map. The depth map offers important information about the objects position in space and helps us estimating their relative distance to the video camera. The proposed method is evaluated on a set of ten 3D video streams and can be considered efficient and robust.