Comments on: Networked Video in 10 Years : Networked Video == Parseable Video http://www.walking-productions.com/notslop/2007/02/26/networked-video-in-10-years-networked-video-parseable-video/ Not - Shawn's Little Outpost Sun, 10 Jan 2010 15:22:14 -0800 http://wordpress.org/?v=2.8.4 hourly 1 By: vanevery http://www.walking-productions.com/notslop/2007/02/26/networked-video-in-10-years-networked-video-parseable-video/comment-page-1/#comment-783 vanevery Tue, 27 Feb 2007 16:13:22 +0000 http://www.walking-productions.com/notslop/2007/02/26/networked-video-in-10-years-networked-video-parseable-video/#comment-783 Hi Dan, I agree, there <i>could</i> be software developed that implements search on video for things like faces and ducks and so on. Unfortunately, every attempt at this that I have seen does a pretty poor job as compared with humans doing the identification. My feelings are that the whole process of producing video needs to change. Instead of creating moving images with a dumb camera that just captures light, moving images should be captured with cameras that know the time and location, that can be told the context and that allow video to be marked up and tagged on the spot. Futhermore, the video shouldn't be one continous run, it should have scene detection, it should record the settings used in software like exposure, white balance, iris, zoom level and so on. This would give us a running start and isn't nessecarily difficult. Hi Dan,

I agree, there could be software developed that implements search on video for things like faces and ducks and so on. Unfortunately, every attempt at this that I have seen does a pretty poor job as compared with humans doing the identification.

My feelings are that the whole process of producing video needs to change. Instead of creating moving images with a dumb camera that just captures light, moving images should be captured with cameras that know the time and location, that can be told the context and that allow video to be marked up and tagged on the spot. Futhermore, the video shouldn’t be one continous run, it should have scene detection, it should record the settings used in software like exposure, white balance, iris, zoom level and so on.

This would give us a running start and isn’t nessecarily difficult.

]]> By: Dan M http://www.walking-productions.com/notslop/2007/02/26/networked-video-in-10-years-networked-video-parseable-video/comment-page-1/#comment-782 Dan M Tue, 27 Feb 2007 15:41:44 +0000 http://www.walking-productions.com/notslop/2007/02/26/networked-video-in-10-years-networked-video-parseable-video/#comment-782 Certainly searchability, which requires 'parseability', and without relying on related/linked text is key to making a really useful video 'encyclopedia' out of the great pile of digital images gathering online. Would not something along the lines of 'facial recognition software' be the direction any such tool would have to take? Obviously the problem is considerably more complex than Western text which can be broken down and completely represented in less than 128 pieces. However I suspect that a study of shapes and their relationships could come up with a reasonable number of checks that would lead to near perfect matches. In other words, if you see a duck as an 'oval with a triangle attached at one end and a curved tube at the other' with a size relationship thrown in you would in fact find way more actual images of ducks than anything else. Allowance would have to be made for ranking 'nearness of match' since while a 't' is a 't' and 'oval' can be many things, but that should be easily doable. A vocabulary of image templates could be built, for example; human faces have certain features within a very limited range of relationships, so do cats, dogs, sailboats, chairs houses, etc., etc., etc. If this scanning and indexing of images was done on an ongoing basis as the search engines do with text, it would seem to me that a routine search for images of this or that could be quite rapid, and at least as accurate as text based searches on a particular subject, without any reliance on linked text clues at all. Certainly searchability, which requires ‘parseability’, and without relying on related/linked text is key to making a really useful video ‘encyclopedia’ out of the great pile of digital images gathering online. Would not something along the lines of ‘facial recognition software’ be the direction any such tool would have to take? Obviously the problem is considerably more complex than Western text which can be broken down and completely represented in less than 128 pieces. However I suspect that a study of shapes and their relationships could come up with a reasonable number of checks that would lead to near perfect matches. In other words, if you see a duck as an ‘oval with a triangle attached at one end and a curved tube at the other’ with a size relationship thrown in you would in fact find way more actual images of ducks than anything else. Allowance would have to be made for ranking ‘nearness of match’ since while a ‘t’ is a ‘t’ and ‘oval’ can be many things, but that should be easily doable.

A vocabulary of image templates could be built, for example; human faces have certain features within a very limited range of relationships, so do cats, dogs, sailboats, chairs houses, etc., etc., etc.

If this scanning and indexing of images was done on an ongoing basis as the search engines do with text, it would seem to me that a routine search for images of this or that could be quite rapid, and at least as accurate as text based searches on a particular subject, without any reliance on linked text clues at all.

]]>