Point-to-point routing is central to communication networks. In this paper, we present a novel addressing and routing scheme for wireless sensor networks. We base our approach on the observation that in real applications, sensors are usually deployed in groups; while it is impractical to predict the landing location for each individual sensor, the locations of sensors from the same group tend to follow certain probabilistic model. By taking advantage of this deployment knowledge, we design a Monte Carlo sampling algorithm that distributedly discovers group-level topology of the sensor field, and represents it as a compact atlas. Meanwhile we assign each node an address comprised of its group ID and local coordinates within the group. Efficient point-to-point routing is achieved as two sub-procedures, proactive path planning on the high-level atlas and reactive actual routing using local coordinates information. In addition, our approach takes account of node density information, and prolongs the network lifetime by conserving the energy of sensors in sparse areas, which is especially important for non-evenly dense sensor networks. Experimental results show that our approach enables efficient and density-aware routing even in environment with complex topology structures.