Visual place recognition is one of the most challenging problems in computer vision, due to the large diversities that real-world places can represent. Recently, visual place recognition has become a key part of loop closure detection and topological localization in long-term mobile robot autonomy. In this work, we build up a novel visual place recognition pipeline composed of a first filtering stage followed by a partial re-ranking process. In the filtering stage, image-wise features are utilized to find a small set of potential places. Afterwards, stable region-wise landmarks are extracted for more accurate matching in the partial re-ranking process. All global and partial image representations are derived from pre-trained Convolutional Neural Networks (CNNs), and the landmarks are extracted by object proposal techniques. Moreover, a new similarity measurement is provided by considering both spatial and scale distribution of landmarks. Compared with current methods only considering scale distribution, the presented similarity measurement can benefit recognition precision and robustness effectively. Experiments with varied viewpoints and environmental conditions demonstrate that the proposed method achieves superior performance against state-of-the-art methods.