[ https://issues.apache.org/jira/browse/SPARK-8486?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Feynman Liang updated SPARK-8486: --------------------------------- Description: Scale invariant feature transform (SIFT) is a scale and rotation invariant method to transform images into matrices describing local features. (Lowe, IJCV 2004, http://www.cs.ubc.ca/~lowe/papers/ijcv04.pdf) We can implement SIFT in Spark ML pipelines as a org.apache.spark.ml.Transformer. Given an image Array[Array[Numeric]], the SIFT transformer should output an ArrayArray[[Numeric]] of the SIFT features for the provided image. The implementation should support computation of SIFT at predefined interest points, every kth pixel, and densely (over all pixels). Furthermore, the implementation should support various approximations for approximating the Laplacian of Gaussian. In addition to approximating using Difference of Gaussian (as described by Lowe), we should support * SURF approximation using box filters (Bay, ECCV 2006, http://www.vision.ee.ethz.ch/~surf/eccv06.pdf) should also be supported. * DAISY was: Scale invariant feature transform (SIFT) is a scale and rotation invariant method to transform images into matrices describing local features. (Lowe, IJCV 2004, http://www.cs.ubc.ca/~lowe/papers/ijcv04.pdf) We can implement SIFT in Spark ML pipelines as a org.apache.spark.ml.Transformer. Given an image Array[Array[Numeric]], the SIFT transformer should output an ArrayArray[[Numeric]] of the SIFT features for the provided image. The implementation should support computation of SIFT at predefined interest points, every kth pixel, and densely (over all pixels). Depending on performance, approximating Laplacian of Gaussian by Difference of Gaussian (traditional SIFT) as described by Lowe can be even further improved using box filters (aka SURF, see Bay, ECCV 2006, http://www.vision.ee.ethz.ch/~surf/eccv06.pdf). > SIFT/SURF Feature Transformer > ----------------------------- > > Key: SPARK-8486 > URL: https://issues.apache.org/jira/browse/SPARK-8486 > Project: Spark > Issue Type: Sub-task > Components: ML > Reporter: Feynman Liang > > Scale invariant feature transform (SIFT) is a scale and rotation invariant > method to transform images into matrices describing local features. (Lowe, > IJCV 2004, http://www.cs.ubc.ca/~lowe/papers/ijcv04.pdf) > We can implement SIFT in Spark ML pipelines as a > org.apache.spark.ml.Transformer. Given an image Array[Array[Numeric]], the > SIFT transformer should output an ArrayArray[[Numeric]] of the SIFT features > for the provided image. > The implementation should support computation of SIFT at predefined interest > points, every kth pixel, and densely (over all pixels). Furthermore, the > implementation should support various approximations for approximating the > Laplacian of Gaussian. In addition to approximating using Difference of > Gaussian (as described by Lowe), we should support > * SURF approximation using box filters (Bay, ECCV 2006, > http://www.vision.ee.ethz.ch/~surf/eccv06.pdf) should also be supported. > * DAISY -- This message was sent by Atlassian JIRA (v6.3.4#6332) --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org