[jira] [Created] (SPARK-8493) Fisher Vector Feature Transformer

["Feynman Liang (JIRA)" <ji...@apache.org>]

Feynman Liang created SPARK-8493:
------------------------------------

             Summary: Fisher Vector Feature Transformer
                 Key: SPARK-8493
                 URL: https://issues.apache.org/jira/browse/SPARK-8493
             Project: Spark
          Issue Type: Sub-task
          Components: ML
            Reporter: Feynman Liang


Fisher vectors provide a vocabulary-based encoding for images (see https://hal.inria.fr/hal-00830491/file/journal.pdf). This representation is useful due to reduced dimensionality, providing regularization as well as increased scalability.

An implementation of FVs in Spark ML should provide a way to both train a GMM vocabulary as well compute Fisher kernel encodings of provided images. The vocabulary trainer can be implemented as a standalone GMM pipeline. The feature transformer can be implemented as a org.apache.spark.ml.UnaryTransformer. It should accept a vocabulary (Array[Array[Double]]) as well as an image (Array[Double]) and produce the Fisher kernel encoding (Array[Double]).

See Enceval (http://www.robots.ox.ac.uk/~vgg/software/enceval_toolkit/) for a reference implementation in MATLAB/C++.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@spark.apache.org
For additional commands, e-mail: issues-help@spark.apache.org