France Telecom Orange Labs (Beijing) AT MSR-BING CHALLENGE ON IMAGE RETRIEVAL 2013

This study addresses approaches of our team ORANGE for MSR-Bing Image Retrieval Challenge to assess the relevance on a pair of query term and image. Our approaches aim to boos the performance of web scale image retrieval, where images on the initial list (indexed by query term) can be re-ranked based on their relevance. One year BING image retrieval logs are used to develop the relevance assessment system. Several visual features are employed to describe one image. A visual similarity learning algorithm is introduced to train a weighted image similarity. Three runs are submitted for evaluation: “boostLearn”, “fast v1”, “learn RF”.