Bag-of-Words Against Nearest-Neighbor Search for Visual Object Retrieval

We compare the Bag-of-Words (BoW) framework with the Approximate Nearest-Neighbor (ANN) based system in the context of visual object retrieval. This comparison is motivated by the implicit connection between these two methods: generally speaking, the BoW framework can be regarded as a quantization-guided ANN voting system. The value of establishing such comparison lies in: first, by comparing with other quantization-free ANN system, the performance loss caused by the quantization error in the BoW framework can be estimated quantitatively. Second, this comparison completely inspects the pros and cons of both ANN and BoW methods, thus to facilitate new algorithm design. In this study, by taking an independent dataset as the reference to validate matches, we design an ANN voting system that outperforms all other methods. Comprehensive and computationally intensive experiments are conducted on two Oxford datasets and two TrecVid instance search datasets, and the new state-of-the-art is achieved.