m24757 - Stanford/Nokia: presented improvements to the global descriptor that had been presented in m23578 of the last meeting. Compared to the previous contribution, memory has been substantially reduced, with a limited drop in accuracy. The proposed global descriptor has 512 Bytes. Significant improvement in performance was achieved compared to TM1. Localization missing at the lowest rate. Further investigation is required on this aspect. Cross-check has been completed for m23578 at last meeting, m24757 will be partially cross-check during the meeting.
m24781 – Peking University: Similarity between images was computed based on Fisher Vectors. GMM models are used for aggregating FVs and PCA for dimensionality reduction. BoW approach extended applying soft assignment. The proposed global descriptor has 256 Bytes, enabling localization at the lowest rate. Accumulated compressed FV based distance is also used for pairwise matching. Significant improvement in performance is achieved in retrieval and matching, but at the cost of very high memory consumption (130MB). Improvements are ongoing to reduce the memory to fit into the 5MB memory limit.
