[
https://issues.apache.org/jira/browse/TIKA-2322?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15967037#comment-15967037
]
Madhav Sharan commented on TIKA-2322:
-------------------------------------
Thanks for your interest Kranthi. It will be great working this out in tika-dl.
Let fix this issue and then we can start investiagting more for adding it in
tika-dl, it will be a good addition there. We need to confirm few things -
- I came to know that OpenCV is pulled internally by DL4J we need to confirm
this and if that's the case it will be super simple for us to use.
- We need to choose best way of using OpenCV I have used below [0], [1], [2]. I
liked [2] the most because of easy integration but last time I used it I faced
lot of memory management issues. We need to double check if we can use these
integrations with tika-server with atleast 1000 videos without server restart.
[0]
https://github.com/USCDataScience/hadoop-pot/blob/master/hadoop-pot-video/pom.xml#L10-L14
[1] https://opencv-java-tutorials.readthedocs.io/en/latest/index.html
[2] https://github.com/bytedeco/javacv
> Video labeling using existing ObjectRecognition
> -----------------------------------------------
>
> Key: TIKA-2322
> URL: https://issues.apache.org/jira/browse/TIKA-2322
> Project: Tika
> Issue Type: Improvement
> Components: parser
> Reporter: Madhav Sharan
> Assignee: Chris A. Mattmann
> Labels: memex
> Fix For: 1.15
>
>
> Currently TIKA supports ObjectRecognition in Images. I am proposing to extend
> this to support videos.
> Idea is -
> 1. Extract frames from video and run IncV3 to get labels for these frames.
> 2. We average confidence scores of same labels for each frame.
> 3. Return results in sorted order of confidence score.
> I am writing code for different modes of frame extractions -
> 1. Extract center image.
> 2. Extract frames after every fixed interval.
> 3. Extract N frames equally divided across video.
> We used this approach in [0]. Code in [1]
> [0] https://github.com/USCDataScience/hadoop-pot
> [1] https://github.com/USCDataScience/video-recognition
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
