The data is there but this needs to be added to the platform
1 Like
Great, the addon works really great and will be my solution atleast until switching to a NUC. Only thing missing now is bounding boxes for me personally. I hope you can find time to add it someday.
Thank you so much for your work.
Do you think it would be more accurate vs a local copy of faster_rcnn_inception_v2_coco?
I am just trying this out since my NUC apparently doesn’t have AVX, that’s needed to run most image processing components locally. Really fast I must say! Is there a list for which targets are available?
HI @teachingbirds I just released v1.0 which will display the list of targets, let me know how you get on with it! Cheers
1 Like
Nice, will install later!
Do you have an example automation showing how you trigger this and handle the response? I have hikvision cameras, and I would like the motion detection to trigger this and then send me a notification if its a person.
I don’t have an automation yet but the state is the number of objects of the target class, so I would have an automation, when state goes from 0 to 1, 2 etc, send an image
Ok, so right now it shows the state as “unknown”. This is the error in the logs.
botocore.errorfactory.InvalidImageFormatException: An error occurred (InvalidImageFormatException) when calling the DetectLabels operation: Request has invalid image format
Connection lost. Reconnecting…
For the source I use the front door camera. Do I need to use a static image instead?
source:
- entity_id: camera.FrontDoor
That is a strange error, this should work with any camera
you need to reset the state to zero before making the scan call.
awesome work mate. tired of compiling tensorflow for every version of python. do we have similar API for google too? let me know if you need to test something. cheers.
You can also setup tensorflow in Google’s cloud platform. I think it’s supposedly cheaper and easier than vision.
https://cloud.google.com/ml-engine/docs/tensorflow/getting-started-training-prediction
Not sure if it’s actually cheaper than vision. I haven’t compared the rates, but I swear I read that someone had and that it was cheaper to do tensorflow. Although I imagine google’s vision is probably more accurate considering who’s maintaining that one.
Also, THANK YOU so much for this. I run in Docker on an old PC with old instruction sets and Tensorflow was a real PITA, I never got it running. This literally took less than 10 minutes. No more false positives from my security cameras. I’d have had it going days ago but apparently I requested to close my AWS account years ago and it took Amazon support 3 days to get it reactivated. (They would only reply once a day and only after I had gone to sleep each day.)
1 Like
I have been running this for couple of weeks now and I noticed the memory usage is kept crawling up. (Other than that It works great).I run my hass on Synology NAS docker. usually it takes about 250mb. But after adding the image_process (this component), it kept going up. roughly at 2gb per day. I left it running for 3 days and it used up 6gb. I wasn’t sure if it was this component was the culprit, but I disabled it this morning and restarted it, and now it is using just 250mb. Do you know of any memory leaks?
Hmm I suggest checking the HA repo for any similar sounding issues
Sorry, I wasn’t clear earlier. My questions was if you knew of any memory leaks in this component. Thanks.
I don’t think so, this behaves as any other image processing component, but please let me know if you conclude otherwise