Visual query compression with embedded transforms on Grassmann manifold

Zhaobin Zhang, Li Li, Zhu Li

Abstract: This paper proposes a novel approach for visual query compression with a bank of transforms selected on Grassmann manifold. Instead of using single transform and quantization pipeline for all the features extracted from an image, we group the key point features according to their local embedding subspace geometry and organize them into different leaf nodes of a KD-Tree which was built from a large dataset. We introduce a new method to find the optimal local transforms on Grassmann manifold such that each feature can achieve the best energy compaction possible given the rate-distortion constraint. Experiments results demonstrated that the proposed method outperforms global transform in both reconstruction error and visual query accuracy, especially in high compression ratio circumstances.