MARQS: RETRIEVING SKETCHES USING DOMAIN- AND STYLE-INDEPENDENT FEATURES LEARNED FROM A SINGLE EXAMPLE USING A DUAL-CLASSIFIER

Authors: Brandon Paulson, Tracy Hammond

Comments:
1. Daniel's blog
2. Ben's Blog

Summary:
MARQS - Media Albums Retrieved by a Query Sketch - an application to which can store music/ pictures and retrieve them with the help of sketch based queries.Applications uses 2 recognizers - simple classifier and a linear classifier (similar to Rubine). The classifiers use set of global features that describe the sketch rather than single stroke. This helps the classifier to be extended to multi stroke sketches.
Sketches are first rotated about an major axis ( line between farthest points on sketch) such that the axis lies horizontal.Features - Bounding box aspect ratio,Pixel Density, Average Curvature and The number of perceived corners across all strokes of the sketch. Simple classifier is used for sketch with one example. It calculates the absolute differences between the feature of sketch and template (Errors), then normalized by dividing them by the value of feature in query sketch.The errors are then summed up and the result is published based on the lowest errors.
Linear classifier (similar to rubine) is applied to sketches with more than 2 samples.

Discussion:
This algorithm is an interesting approach. it is not affected by the direction of stroke/ scale/ change in local orientation. (None of the features depend on these factors). I do not think pixel density is an important feature- I would have to look into how this feature affects sketch recognition. Anyways it would be taken care by the weights calculated by the linear classifier.
I would like to know if there are features which can fit in this algorithm.