Fried Binary Embedding: From High-Dimensional Visual Features to High-Dimensional Binary Codes.

Area of Science:

• Computer Vision
• Machine Learning
• Data Science

Background:

• Existing binary embedding methods often use compact codes, limiting the use of discriminative high-dimensional visual features.
• Generating long binary codes is computationally and memory intensive due to large projection matrices.

Purpose of the Study:

• To propose efficient methods for generating long binary codes from high-dimensional visual features.
• To reduce the computational and memory complexity of binary embedding.
• To enhance accuracy in tasks like approximate nearest neighbor search.

Main Methods:

• Developed Fried binary embedding (FBE) for unsupervised scenarios and Supervised Fried Binary Embedding (SuFBE) for supervised scenarios.
• Utilized adaptive Fastfood transform to decompose the projection matrix into structured matrices.
• Reduced computational complexity from O(d*k) to O(d*log k) and memory cost from O(d*k) to O(d*log k).

Main Results:

• FBE and SuFBE significantly decrease computational and memory requirements compared to traditional methods.
• The structured matrices in FBE and SuFBE regulate the projection matrix, leading to improved accuracy.
• Experimental results show superior efficiency and performance over state-of-the-art methods on various visual applications.

Conclusions:

• FBE and SuFBE offer an efficient and effective approach to generating long binary codes.
• These methods enable better utilization of high-dimensional visual features for improved performance in computer vision tasks.
• The proposed techniques provide a practical solution for memory and compute-intensive binary embedding challenges.