Improving bag-of-words representation with efficient twin feature integration

In recent years, the Bag-of-Words (BoW) model has been widely used in most state-of-the-art large-scale image retrieval systems. However, the standard BoW based systems suffer from low discriminative power of local features as well as quantization errors that significantly affect the retrieval performance. In this paper, twin feature is employed and well combined with two advanced techniques including Hamming Embedding (HE) and Multiple Assignment (MA) to construct a discriminative image retrieval system on BoW representation in an efficient way. Experimental results on two benchmark datasets Oxford5k and Paris6k demonstrate that the proposed technique can greatly refine the visual matching process and enhance the final performance for image retrieval.