We propose a novel 3D tracking method that supports several hundreds of pre-trained potential planar targets without losing real-time performance. This goes well beyond the state-of-the-art, and to reach this level of performances, two threads run in parallel: the foreground thread tracks feature points from frame-to-frame to ensure real-time performances, while a background thread aims at recognizing the visible targets and estimating their poses. The latter relies on a coarse-to-fine approach: assuming that one target is visible at a time, which is reasonable for digilog books applications, it first recognizes the visible target with an image retrieval algorithm, then matches feature points between the target and the input image to estimate the target pose. This background thread is more demanding than the foreground one, and is therefore several times slower. We therefore propose a simple but effective mechanism for the background thread to communicate its results to the foreground thread without lag. Our implementation runs at more than 125 frames per second, with 314 potential planar targets. Its applicability is demonstrated with an Augmented Reality book application.