Pull request for TIKA-1798 Parser for Video Similarity using PooledTimeSeries metric contributed by Aditya Dhulipala and Chris Mattmann.

[chrismattmann]

No description provided.

[chrismattmann]

needs to be generalized.

[chrismattmann]

ok @cafed00d4j I significantly cleaned this up and got it working! Great initial work!

Some comments:

1. I thought we discussed using a table here. Since you have # of rows and # of cols present in the div, why not simply use tables and columns here to represent the OF and OG? (I'll fix this now).
2. you don't have an ending div on your XHTML (I'll add a fix for this now).
3. You don't identify in the XHTML what is the OF and what is the OG (I'll fix this now).
4. We should probably start out with an h3 of the file name to start out the display.
5. I get a whole bunch of NaNs before the XHTML output. Not sure where those are from but they are printed to stdout.

[adi-dhulipala]

Hi Prof. Chirs,

Thanks for the feedback.. Please check my responses below.

1. I thought we discussed using a table here. Since you have # of rows and # of cols present in the div, why not simply use tables and columns here to represent the OF and OG? (I'll fix this now).
   ------>
   In one of the progress reports I think you mentioned we can use "ol li --- li ol" sequences instead
   The OF and HOG are series of 1D vectors instead of series of 2D matrices as I originally thought.
2. you don't have an ending div on your XHTML (I'll add a fix for this now).
3. You don't identify in the XHTML what is the OF and what is the OG (I'll fix this now).
   ------>
   Sorry about that. I think I missed it
4. We should probably start out with an h3 of the file name to start out the display.
   ------>
   Understood. Will fix it now
5. I get a whole bunch of NaNs before the XHTML output. Not sure where those are from but they are printed to stdout.
   ------>
   Yes. These are the meanChiSquare distance measures of the FeatureVectors. Since, we're running this on a single file, we only have one set of FeatureVectors.
   When PoT is run on a collection of video files, we extract each video's OF and HOG and then calculate the meanSquareDist across all of them.
   But this may not be within the scope of a Tika parser. Maybe it should be included in the wrapper program that calls Tika on a set of videos. Much like the implementation of tika-img-similarity github project. Do you think that's right?

[chrismattmann]

@cafed00d4j i don't get the response to 1.i. - if OF and HOG Are 1D vectors why is there a num rows, num cols?

[adi-dhulipala]

I couldn't think of better terms for them.

The numRows is number of time-series in the video. This proportional to the numFrames in the video. But it's not equal to it. It depends on the how the underlying computeFarnebackFlow() and computeGradeient() operations behave. but the general term the paper uses is time-series and each of the rows is a unit of time-series

The numCols is a constant. Of size 200. It's the vector size. It's got by applying some pooling operators on the flowVectors and gradientHistorgrams. (It's initialized around line 412. There's also a comment there about the chosen size)

So, each row in the output is a PoT vector of that particular time-serie.

[chrismattmann]

Thanks @cafed00d4j so I think we can still use a table here and just do rowspan to take care of non uniform sizes - colspan isn't needed since it's always 200 right?

[adi-dhulipala]

Hi Prof.

No problem!

Yes. I think that can omitted. But we can leave it in for completeness it's that's important.