Vocabulary tree-based method is one of the most popular methods for content-based image retrieval due to its efficiency and effectiveness. However, for existing vocabulary tree methods, the retrieval precision in large scale image database has never been acceptable especially for image datasets with high variations. In this paper, we propose a novel tree fusion framework: Feature Forest, utilizing and fusing different kind of local visual descriptors to achieve a better retrieval performance. In the offline-learning stage, our framework first establishes different feature vocabulary trees based on different features and uses the average covariance to build vocabulary tree adaptively. In the online-query stage, we use the ratio of the resulting score to the standard score to fuse retrieval results of each vocabulary tree adaptively. The evaluations show the effectiveness of our approach compared with single vocabulary-tree based methods on different databases.