Open wizu opened 10 months ago
Yep. This is already planned.
Ooh, I just thought of an extension to this, when labeling is part of local Frigate: when you draw a new bounding box, re-run it through object detection, and automatically select a label for it based on the most likely object, even if it's a very low score on the current model that wasn't high enough to trigger a box in the first place. This comes because I was just labeling a bunch of images, and drew a bounding box around a person, and was going quickly and didn't realize the selection was still on car
when I hit save.
For what it's worth, I've largely stopped verifying images in Frigate+ until this is resolved. I'm still uploading images, it just takes too much time to add bounding boxes for all the other objects (to hundreds of images) at the moment. Will it be possible to retroactively run detection over all my unverified images? Is there anything we in the community can do to assist with this issue, given I don't think the Plus server is open source?
I'm also stopping until this solution is released, it takes me too much time every day if I want to add labels to each object on my camera... I have up to 10 labels per image at all times.
Will 0.14 be the release of this feature?
Just jumping in to keep notified when this changes status. A very helpful feature for sure!
Yep. This is already planned.
For 0.14 is it possible? frigate+ would be of much more interest.
My driveway has neighbors cars visible most of the time + my car + license plates. So lots of stuff to add on each uploaded image. I agree that we should be able to set a treshold value in the config, and it would upload all objects above the treshold. I am considering ignoring my neighbors cars zone, as they become difficult to detect at night.
How was this not done OOTB?
+1, glad you're already planning it. Too much effort for each image today, will probably occasionally mark false positives for now and wait for this to consider Frigate+.
Describe what you are trying to accomplish and why in non technical terms It would be great, if object detection could be run again before sending snapshot to Frigate+ and all objects would have bounding boxes and labels already set. Then we could tweak them to perfect size and set correct labels. My main camera watches my parking, where two of my cars parks. There is also often my dog. So almost every time i send image to Frigate+ to label it, I have to mark two cars, two license plates and a dog. It's time consuming :/
Describe the solution you'd like When a snapshot is sent do Frigate+, object detection should be run without any label exclusion, and found objects should be send along with snapshot.
Describe alternatives you've considered Object detection could also be run on Frigate+ server, with usage of main detection model.
Additional context In many of my snapshop I have to mark more than a 5 of labels.