A Method of Memory Reduction for Specific Object Recognition with a Bloomier Filter

Specific object recognition based on nearest neighbor search of feature vectors requires a huge amount of memory to store all feature vectors for distance calculation. To solve this problem, we propose a memory reduction method for specific object recognition with no distance calculation for matching feature vectors. The proposed method is characterized by the use of a Bloomier filter, which is far memory efficient than a hash table, for the storage of feature vectors. The proposed method is evaluated based on experiments of planar and 3D specific object recognition in comparison to a method with a hash table.