New Objects for alerts / too many? (and a request)

[Robert Spendlove]

I see a lot of requests for new objects for alerts and begin to wonder if there are too many.  Does it help my processor load to only look for a few specific types of objects vs all of them?   As an extreme example, if I have "person" exclusive on all cameras, is there less processor being used than if it looks for any object?    Similarly, does having small detection areas also help this?  

I originally though it was odd that there were so many small animal categories like raccoon, rabbit, mouse and skunk when "small animal" or "unknown small animal" would suffice, but deer wasn't detected.  I know that deer has been added and it's a good addition but I still would not add elk, because it looks so similar to a deer to the camera and AI... like a wolf and dog are similar.    Just lump them together, even if it needs to be renamed "deer/elk".

I don't know about the robot mower.... the general shape is similar to a car or van, but smaller.  Again, the AI doesn't seem to know scale very well, as it used to detect a West Highland White Terrier as a (polar) bear sometimes.    Perhaps just have the system NOT alert to cars in the grassy area?  Tell the system that there is one of these at the property and it can try to differentiate for alerts, maybe watching it longer before sending an alert?  Just throwing out some ideas here.

One addition I'd like to request is for a boat.  I have seen video from someone that had their camera looking over a dock area and they had lots of animals detected, but I'm sure they'd like to be notified if a boat entered their dock.    I do have a boat on the side of the house, and see it in the road when I pull it out so sure... I'd like to know if it moves through the driveway but it's a rare use case for me. This would be more useful for the lucky saps that bought a house or cabin on lakefront property.  I can provide footage of it on a trailer if it helps, but I wouldn't want the AI to always look for the trailer wheels to detect a boat.

[CamectChao]

On Tuesday, May 12, 2020 at 2:45:38 PM UTC-7, Robert Spendlove wrote:

I see a lot of requests for new objects for alerts and begin to wonder if there are too many.  Does it help my processor load to only look for a few specific types of objects vs all of them?   As an extreme example, if I have "person" exclusive on all cameras, is there less processor being used than if it looks for any object?   

No, not with the current implementation. 

Technically, it could help a bit if the number of objects is smaller, very tiny though. We don't do such optimization because it'll make the related code much more complex, not worth the slightly performance improvement.

 

Similarly, does having small detection areas also help this?  

It depends. If you point a camera at a busy street, creating a negative area (w/o any object specified) covering the street does help a lot. In this special case, we won't run AI if there is any motion on the street.

I originally though it was odd that there were so many small animal categories like raccoon, rabbit, mouse and skunk when "small animal" or "unknown small animal" would suffice, but deer wasn't detected.  I know that deer has been added and it's a good addition but I still would not add elk, because it looks so similar to a deer to the camera and AI... like a wolf and dog are similar.    Just lump them together, even if it needs to be renamed "deer/elk".

Right. With the video quality from today's security camera, it's unlikely the AI could tell the difference between deer and elk.

I don't know about the robot mower.... the general shape is similar to a car or van, but smaller.  Again, the AI doesn't seem to know scale very well, as it used to detect a West Highland White Terrier as a (polar) bear sometimes.    Perhaps just have the system NOT alert to cars in the grassy area?  Tell the system that there is one of these at the property and it can try to differentiate for alerts, maybe watching it longer before sending an alert?  Just throwing out some ideas here.

Right. We are trying to teach the AI to ignore these robots instead of recognize it as a separate object. 

One addition I'd like to request is for a boat.  I have seen video from someone that had their camera looking over a dock area and they had lots of animals detected, but I'm sure they'd like to be notified if a boat entered their dock.    I do have a boat on the side of the house, and see it in the road when I pull it out so sure... I'd like to know if it moves through the driveway but it's a rare use case for me. This would be more useful for the lucky saps that bought a house or cabin on lakefront property.  I can provide footage of it on a trailer if it helps, but I wouldn't want the AI to always look for the trailer wheels to detect a boat.

Yeah. Detecting boat on a trailer might not be a popular use case, and is slightly more difficult than detecting boat on the water.

We may add the later if many people want it...

[Jack 7]

Does Camect use the camea mainsteam or the camera substream to do motion detection?

[Eric Meeson]

I believe it uses the main stream as default because that's what auto-discovery finds. If you manually set the substream in it, it will use that, but you'll notice that your video resolution is much lower.

You can check for sure by clicking the wrench on your camera, then clicking the "i" next to camera name and looking at the size.

I get detections from far enough away I don't see how it could be using the substream, the image would be too small/blurry if it was.

[Jack 7]

I realize that mainstream is normally recorded but was wondering if mainstream is constantly decoded for motion/object detection. That would normally take a lot of processing power. I noticed recently that Blue Iris started using substream for motion detection, producing large reductions in cpu utilization. I was hoping a Camect developer might comment.

[CamectArup]

Everything is done from one stream, which is the main stream by default. If you want to use a substream, you can add your camera manually using the RTSP url for the substream, and disable the auto-detected copy of the camera that's pulling the main stream. However, it will then be the substream that's used for everything. 

We actually do recommend using a substream to people that have a large number of 4K cameras on a regular NVR ... For those cases, you can usually get perfectly good results recording a 1080p substream, allowing you to put a larger number of cameras onto Camect than you would if also using it to record the 4K video stream.