Martin Thoma

## MathBrush The data is on another scale. The datapoints are spread much more. This makes it necessary to implement scaling for data visualization and segmentation.

Figure 4 of "Mathematical expression recognition: a survey" shows that the structural analysis cannot be done locally

Segmentation by strokes seems to be possible for mathematical context. Be aware that the order of strokes is not necessarily ordered (e.g. enlarging fraction strokes)

For http://www.martin-thoma.de/write-math/view/?raw_data_id=151305 the classifier predicted the segmentation `[[0,4],[1,2,5,6],[3]]`:  The yellow and black segmentation does not make any sense, as it would mean there is a symbol between strokes of...

## 153892 Real segmentation: `[[0, 2, 16], [1], [3, 4], [5, 6], [7, 8], [9, 10, 11], [12, 13], [14, 15]]` (got at place -1) Predict segmentation: `[[0, 4, 6],...

``` [157, 2, 498, 59, 14, 33, 4, 4, 3, 1, 4, 14, 13, 0, 0, 2, 11, 277, 1, 1, 1, 1, 5, 4, 74, 1, 1, 1, 4,...

Idea: ## Analyze data Search single-stroke symbols which are not "prefixes" of other symbols. For example. `o` is a prefix of `\oplus` or `l` is a prefix of `t`; `1`...

Idea: Use MST as described in http://www.ai.mit.edu/projects/natural-log/papers/matsakis-MEng-99.pdf#page=37 - Each node is the center of the bounding box of a stroke. - A segmentation of the strokes equals a segmentation of...

Ideas: - Go through each stroke, classify it. If the net is sure enough (gives a probability over a threshold, e.g. > 95%), the recording will be segmented there -...

Other problems: - Wrong segmentation: `\doteq` vs `\stackrel{!}{=}` - Wrong segmentation for "H": http://www.martin-thoma.de/write-math/view/?raw_data_id=324179