220 likes | 424 Views
SURF Feature Compression. EE398 Winter 2008 Ivan Janatra June Zhang. Outline. SURF Features Transform coding Quantization Encoding Results. SURF Features. Given a set of SURF feature descriptors. Explore how compression affect feature descriptors and their performance. Query image.
E N D
SURF Feature Compression EE398 Winter 2008 Ivan Janatra June Zhang
Outline • SURF Features • Transform coding • Quantization • Encoding • Results
SURF Features • Given a set of SURF feature descriptors. Explore how compression affect feature descriptors and their performance Query image Database image
feature descriptor feature descriptor 1x64 1x64 SURF Features database query
Database feature descriptor query feature descriptors SURF Features In 64D feature space MATCH, if Distance 1 < 0.8 * Distance2 Distance 2 Distance 1
Database feature descriptor query feature descriptors SURF Features In 64D feature space Distance 2 Distance 1
SURF Features Pre RANSAC feature matching
SURF Features Post RANSAC feature matching
orthonormal transform quantization Lossless encoding query feature descriptors de- quantization Lossless decoding reconstructed query feature descriptors inverse orthonormal transform Match features database feature descriptors Codec Structure MSE Match rate
Transform Coding: KLT Coding gain= = 3.9
Transform Coding: DCT Coding Gain = = 1.31
Codec Structure KLT quantization Lossless encoding query feature descriptors transformed coefficients symbols Lossless decoding bit stream
Quantization • Mid-tread uniform quantizer • Optimal for MSE distortion and high rates Step=2^bit
Lossless encoding • KLT guarantees statistical independence of the transformed descriptor coefficients • Transformed descriptor coefficients can be coded separately without loss in efficiency • Symbol statistics generated from database feature descriptors
Lossless encoding: Huffman • Generate a Huffman code table for each transformed descriptor coefficients • Wasteful for low energy coefficients
Lossless Encoding: Run length+ Huffman • Low energy coefficients tend to have long run length • Use run length encoding symbol has a probability of occurrences higher then 0.7
Lossless Encoding: Binary Arithmetic Coding • actually K-ary encoding but binary encoding is sufficient
Conclusion • Typical image based transform coding does not work well for feature descriptors • Arithmetic coding performs close to Shannon Lower Bound • Nearly linear relationship between MSE and match rate