Supervised Multi-scale Locality Sensitive Hashing
read more
Citations
Optimization of deep convolutional neural network for large scale image retrieval
Nonlinear Discrete Hashing
Deep learning hashing for mobile visual search
Privacy-Preserving Outsourced Media Search
Nonlinear Sparse Hashing
References
Distinctive Image Features from Scale-Invariant Keypoints
Modeling the Shape of the Scene: A Holistic Representation of the Spatial Envelope
Locality-sensitive hashing scheme based on p-stable distributions
Spectral Hashing
Related Papers (5)
Frequently Asked Questions (13)
Q2. What is the notion of bit reliability used for estimating the quality of each bit?
The notion of bit reliability is used for estimating the quality of each bit, which is defined as a weighted average of the false positive rate and the false negative rate in a hypothesis test during the training stage.
Q3. What is the main advantage of the proposal?
The main advantage of the proposal is its versatility and thus the potential to be applied to other feature-independent hash algorithms.
Q4. What scale is the for a feature-independent algorithm?
For each scale index si = s0 + i (i = 0, 1, · · · , x− 1), the authors evenly divide the feature vector into l = 2si parts and compute ko (ko ≥ k) hash bits from each part, so that n′ = l · ko.
Q5. How do the authors compute a feature vector?
In order to hash (project) a feature vector, there are at least two ways: the authors could either compute l × k bits from the whole feature vector, or divide it into l parts and compute k hash bits from each part.
Q6. What is the main idea of the work?
In this work, the authors improve the classic LSH framework by incorporating two novel elements: supervised hash bit selection and multi-scale feature representation.
Q7. What is the way to solve the problem?
Since LSH can support arbitrary hash lengths, their solution to the above problem is the following:• Given an LSH algorithm, generate no-bit output (no ≥ n), form a hash value with improved performance by selecting n bits out of no bits.
Q8. What is the reliability of a hash bit?
The reliability of a hash bit can be characterized by the false positive rate pfp and the false negative rate pfn:• pfp = Probability {di = 0
Q9. What are the two parameters that are important?
The two parameters l and k are important - the former typically corresponds to the number of hash tables; the latter is the size of a sub-hash value.
Q10. How much is the cost of a pseudo-random number generator?
In the worst case, when all the candidate hyperplanes are generated online by a pseudo-random number generator and are all used for projection (despite that not all results are used), the cost is about x · k0k times the cost of the original LSH.
Q11. What is the ROC curve for the two scenarios?
The hash size (64, 128, 256); • The size of the bit selection pool (200%, 400%); • The number of available feature scales (1–3);Hypothesis testing is used for evaluating SMLSH in both scenarios.
Q12. How many scales can be used to select a bit?
64 128 256 k 128,256 256,512 512,1024 l 1 x 1 2 3 1 2 3 1 2 30 500 1000 1500 2000 2500 3000 3500 4000 0.440.450.460.470.480.490.5indexbi t rel iabi litybit selection according to reliabilitybit (1st scale) bit (2nd scale) bit (3rd scale) bit (4th scale) selected bit (6.25%)Figure 4: Bit selection from four scales.
Q13. What is the ROC curve for a hash?
In Fig. 5a-b, when the number of scales increases from one to two, the ROC curves typically intersect at a certain middle point: on the left the performance is decreased and on the right the performance is increased.