A saliency map provides information about the regions inside some visual content (image, video, ...) at which a human observer will spontaneously look at. For saliency maps computation, current research studies consider the uncompressed (pixel) representation of the visual content and extract various types of information (intensity, color, orientation, motion energy) which are then fusioned. This paper goes one step further and computes the saliency map directly from the MPEG-4 AVC stream syntax elements with minimal decoding operations. In this respect, an a-priori in-depth study on the MPEG-4 AVC syntax elements is first carried out so as to identify the entities appealing the visual attention. Secondly, the MPEG-4 AVC reference software is completed with software tools allowing the parsing of these elements and their subsequent usage in objective benchmarking experiments. This way, it is demonstrated that an MPEG-4 saliency map can be given by a combination of static saliency and motion maps. This saliency map is experimentally validated under a robust watermarking framework. When included in an m-QIM (multiple symbols Quantization Index Modulation) insertion method, PSNR average gains of 2.43 dB, 2.15dB, and 2.37 dB are obtained for data payload of 10, 20 and 30 watermarked blocks per I frame, i.e. about 30, 60, and 90 bits/second, respectively. These quantitative results are obtained out of processing 2 hours of heterogeneous video content.