Human detection algorithm needs serious work....

[sorka]

It's obvious that the algorithm is using an encompassing boxed ratio to detect the approximate proportions of a human in a rectangle, but this is rediculous:

https://picasaweb.google.com/lh/photo/P ... directlink

A complete fail on what should have been an obvious human detection. Note the green wide box because it was including my shadow so it never detected me as human. I know the current algorithm is very gentle on CPU usage. My current 5 year old I5 2500s PC is only using 6% of the CPU with 9 cameras going. I think we can afford an algorithm that is far more computationally aggressive.

[sorka]

Another odd example. This one just kept firing off over and over thinking my blank wall was a person:

https://picasaweb.google.com/lh/photo/I ... directlink

And this one where it thinks my dog is a person even though the ratio doesn't seem to come close to matching:

https://picasaweb.google.com/lh/photo/_ ... directlink

[Devin]

Hi Sorka,

I'm sorry to hear you've experienced these false detections. We have a dedicated team that is constantly working to improve our algorithms. There are a huge amount of variables which can cause false detections and we are doing our best to address these variables and make Sighthound and as accurate as possible. Please see this article for some general tips for getting the best recognition: https://www.sighthound.com/support/tips ... ecognition.

It's hard to tell because you were filming the screen, but the last two videos you posted looked as though they had low frame rates which will cause issues with the recognition. The recognition is most accurate with at least 10 fps.

Regards,

-Devin

[sorka]

Intresting. The nighttime one indeed has a very low frame rate because it's doing the usual integration of multiple frames to make a useable image.

I can probably fix the second one. My primary video surveillance system is hooked up to all the cameras because of the advanced, albeit none object, motion detection. I think that particular camera can support streaming video so I'll see if I can switch over to that.

The first video is an example of what happens a lot whenever there's any sun casting a shadow left or right.

For the night time camera in the example, I think I can lower the integration and the resolution which should boost the frame rate rate.