Spatial Coding for Large-scale Partial-duplicate Image Search

Опубликовано 28 июля 2016, 22:48

Bag-of-visual-words model is widely used in the state-of-the-art large-scale image retrieval system. It represents each image as a bag of visual words by quantizing local image descriptors to the closest visual words. However, feature quantization reduces the discriminative power of local features, which causes many false visual word matches. Recently, some geometric verification methods are proposed to check the geometric consistency of matched features in a post-processing step. Although retrieval precision is improved, either the computational cost is too expensive to ensure real-time response, or they are limited to local verification. To address this dilemma, we propose a novel scheme, Spatial Coding, designed for large scale partial-duplicate image retrieval. The spatial relationships among visual words are encoded in global region maps. Based on the region maps, a spatial verification approach is developed, which can detect false matches of local features efficiently, and consequently improve retrieval performance greatly. Experiments in partial-duplicate image retrieval, using a database of one million images from Image-Net, reveal that our approach can effectively detect duplicate images with rotation, scale changes, occlusion, and background clutter with very low computational cost. The spatial coding achieve an 53% improvement in mean average precision and 46% reduction in time cost over the baseline Bag-of-Visual-Words approach, respectively. They perform even better than full geometric verification while being much less computationally expensive. Our demo on 10-million dataset further reveals the scalability of our approach.