Keywords:

• handwritten
• paleography
• fragments
• SIFT
• sparse coding
• dictionaries

Q1: What is the ultimate goal of the authors?

The two main goals are providing tools to bring together fragments of the same page (specifically from the Cairo Genizah) and trying to classify handwriting and dates.

Q2: How is SIFT used?

SIFT is used at various points of a letter to generate descriptors. There are 100,000 descriptors overall before inputting them into k-means. SIFT serves as the main classification technique.

Q3: How did they produce the letter dictionaries?

They produced letter dictionaries using the generated SIFT descriptors and k-means clustering to find representative visual words.

Q4: What is sparse coding and its importance?

Sparse coding is used to code documents (or any other codable thing) with a separating/descriptive code which also shows the similarity between items. An example might be the bag of visual words approach.

Q5: Are there any relevant techniques for our research?

This is relevant to the Divan matching problem. This work might be cited in historical document matching, although it covers techniques that are generally well-known in the field.

Keywords:

• LMS estimator
• O(n logn)
• bracelet
• slab
• random
• approximation
• quantiles

Q1: What is LMS?

Given a set of points $p_0, …, p_n$, LMS finds a line $q_0, q_1$ that minimizes the median of the square of distances of $p_0, …, p_n$. This is in contrast with summing up all the squared distances and minimizing them as in OLS (Ordinary Least Squares).

Q2: How are approximations done using LMS?

There are exact solutions for the LMS problem. This paper presents an algorithm with lower complexity. It tries to find an LMS approximation in a band defined by a parameter $\epsilon_r$.

Q3: On which parameters does the algorithm depend?

The algorithm approxLMS depends on a set of lines, a set of quantiles, and two error bounds $\epsilon_r$ and $\epsilon_q$.

Q4: How can this help in our line approximations?

This paper is aimed towards providing an efficient algorithm for random approximations. Since we need repeatability in our keypoint detection, it might be harder (or impossible) to prove that the algorithm produces the exact same line endings in each run with a similar set of points. Hence, it’s not usable in our studies.