International Conference on
Computer Vision 2017
Hashing with Binary Matrix PursuitFatih Cakir†, Kun He‡, Stan Sclaroff‡
†FirstFuel, [email protected]‡Computer Science, Boston University, {hekun,sclaroff}@cs.bu.edu
ExperimentsSummary Formulation
Main contribution:
Technical and empirical improvements for two-stage hashing methods.
Two-stage hashing breaks the problem into two stages:
i. Binary embedding/code inference (affinity matching task)
b
Hamming space
Affinity matrix
ii. Hash function learning (binary classification tasks)
CLASSIFIER
Target binary code
Hashing aims to learn binary embeddings while preserving
the “neighborhood structure” of the data:Input
spaceHamming
space
Neighborhood structure is generally defined via an affinity
matrix.
We propose:
1. Theoretical properties for the binary code inference stage.
2. How to construct the affinity matrix.3. Given insights from (1) - (2), the novel hashing
method “Hashing as Binary Matrix Pursuit” that achieves SOA results on retrieval benchmarks.
Ordinary Hamming distances are unable to reconstruct
certain affinity matrices.
With the projected gradient method the norm of the
residual monotonically decreases.
If non-orthogonal directions are selected at each
iteration then the residual strictly decreases.
CIFAR-10, NUSWIDE, ImageNet100, LabelMe benchmarks
Norm of the residual matrix vs. iteration, when step sizes in the projected gradient descent are constant ( ) and adaptive ( ), corresponding to ordinary and weighted Hamming distances, respectively.
ii. Image retrieval experiments with Hamming rankings
i. Reconstructing the affinity matrix with ordinary and weighted
Hamming distances
• 5K/1K train and test split
• 50K/10K train and test split
https://github.com/fcakir/