Recognizing text through sound alone.

Paper
Li, Wenzhe, and Tracy Anne Hammond. "Recognizing text through sound alone." Twenty-Fifth AAAI Conference on Artificial Intelligence. 2011.
Direct Link: http://www.aaai.org/ocs/index.php/AAAI/AAAI11/paper/download/3791/4119

Summary

This paper introduces a novel approach to sketch recognition using the sound profile of a sketch drawn by scratching on a surface. The authors combine time domain features (mean amplitude) with frequency domain features (Mel-Frequency cepstral coefficients) from the sound profile, after pre-processing to achieve 80% recognition accuracy on recognizing letters in the alphabet sketched out in a constrained manner.

Discussion

Pros

The work is quite novel. I believe GoogleX came up with an idea that is somewhat similar to this one, using similar properties of surface interaction as input for android.

Cons.

The input is constrained to a given surface and a given user.

The input is also constrained to letters drawn out a specific way.

Some of these constraints can be overcome. The authors did not provide metrics on performance without these constraints so the reader can get a sense of how much improvement was gained as a result of so many constraints.

PaleoSketch: accurate primitive sketch recognition and beautification

Paper
Paulson, Brandon, and Tracy Hammond. "PaleoSketch: accurate primitive sketch recognition and beautification." Proceedings of the 13th international conference on Intelligent user interfaces. ACM, 2008.
Publication Link:http://dl.acm.org/citation.cfm?id=1378775

Summary

The work presents a primitive sketch recognition and beautification system known as paleosketch. The idea behind paleosketch is to recognize sketches based on a bottom up approach of identifying low-level primitive shapes as components that combine to form a recognizable high-level shape. The second stage of this system is to return a beautified version of the recognized shape.
To achieve this, they develop two new features in the pre-recognition stage: the normalized distance between direction extremes (NDDE) and the direction change ratio (DCR). The former computes the the difference between the point of highest direction value (ie dy/dx) and the lowest value normalized by stroke length.This feature is able to identify curved shapes (high NDDE values) from poly-lines which have lower NDDE values. The latter DCR value is computed as the maximum change in direction divided by the average change. This value is higher for a poly-line, whereas curves have a much lower value in comparison.

Discussion

Pros

The work is very thorough in presenting the details involved in the implementation.
They introduce two novel features for sketch recognition.

Cons.

A lot of thresholds are used, which are based on training data. Which seems like a lot of tuning. I would like to see how their results change as these different parameters are adjusted.

Combining corners from multiple segmenters.

Paper
Wolin, Aaron, Martin Field, and Tracy Hammond. "Combining corners from multiple segmenters." Proceedings of the Eighth Eurographics Symposium on Sketch-Based Interfaces and Modeling. ACM, 2011.
Publication Link: http://dl.acm.org/citation.cfm?id=2021185

Summary

This work proposes a very interesting meta-classifier for corner finding. By treating the outputs of an ensemble of corner-finding classifiers as features (candidates for corners in a given sketch), the authors develop a corner subset selection process using a weighted error metric to select the best subset of features (which ideally should be the corners contained in the sketch).
They test and compare the performance of their classifier with existing methods to show an improvement in all-or-nothing accuracy.


Discussion

Pros

I like the approach of treating corner candidates as features and selecting the best set.

Cons.

I wonder how this system will work with n-1 corner finding classifiers. I either missed that or it isn't included in the work. The authors motivate the advantage of each classifier, but can they perform just as well or better (or worse) with one less classifier?

Revisiting ShortStraw: improving corner finding in sketch-based interfaces

Paper
Xiong, Yiyan, and Joseph J. LaViola Jr. "Revisiting ShortStraw: improving corner finding in sketch-based interfaces." Proceedings of the 6th Eurographics Symposium on Sketch-Based Interfaces and Modeling. ACM, 2009.
Publication Link: http://dl.acm.org/citation.cfm?id=1572759

Summary

This paper proposes a set of improvements to a previous corner finding algorithm titled shortstraw. In this work, the authors address the following shortcomings from the previous version of shortstraw:
1. Rigid resampling rate
2. The occurrence and consequences of a false corner. The authors address this by using a two staged corner finding stage with aggressive and relaxed threshold values in respective order.
3. The use of a dynamic threshold based on length of line segment to avoid missed corners
4. Sharp noise resulting from resampling: They device an elegant solution using the the behavior of angles around a point on an arch vs a corner to discern the difference between the two.


Discussion

Pros

Addresses much of the shortcomings of the shortstraw paper.
The work is well presented. Their solution to sharpnoise is very elegant

Cons.

It makes the shortstraw solution a little more complicated to implement.
I think their results should have focused more on the specific errors they were trying to address and not the overall performance metrics. Example, isolate cases where shortstaw falsely detects a corner for reason x from the improvement list. Then show how many of those are correctly resolved in the new algorithm.