Many scenarios require a robot to be able to explore its 3D environment online without human supervision. This is especially relevant for inspection tasks and search and rescue missions. To solve this high-dimensional path planning problem, sampling-based exploration algorithms have proven successful. However, these do not necessarily scale well to larger environments or spaces with narrow openings. This paper presents a 3D exploration planner based on the principles of Next-Best Views (NBVs). In this approach, a Micro-Aerial Vehicle (MAV) equipped with a limited field-of-view depth sensor randomly samples its configuration space to find promising future viewpoints. In order to obtain high sampling efficiency, our planner maintains and uses a history of visited places, and locally optimizes the robot's orientation with respect to unobserved space. We evaluate our method in several simulated scenarios, and compare it against a state-of-the-art exploration algorithm. The experiments show substantial improvements in exploration time ($2\times$ faster), computation time, and path length, and advantages in handling difficult situations such as escaping dead-ends (up to $20\times$ faster). Finally, we validate the on-line capability of our algorithm on a computational constrained real world MAV.