I would like to be able to handle the scenario where I have n-classes all with the same number/kind of keypoints. I’m not sure if this is currently possible the way classes are setup in pose. For example, I have 11 keypoints for a race car, and I want to classify the driver for each race car
Hi there!
I would do this in a two stage model workflow. First run a keypoint detection model on the full image, then run a classification model on the driver. You can get the driver with another driver object detection model, or use the keypoint detection bounding box to crop!
Can you tell me more about your keypoint use case and how points might be used when doing a driver classification?
To clarify, what I’m really classifying is the race car (ie same bounding box as keypoints), it’s just easiest to name the class the driver’s name. since the output data is setup like class_number, x,y,w,h,k1x,k1y,k1v,… i figure all I’m really changing here is the class_number from the existing setup. I can’t do two models because this is realtime on edge device and needs to be done in a single pass
How I envision this in roboflow would be grouping the classes. All grouped classes share the same keypoints.
Right, there is no way to group classes for keypoints currently. I can see how that would be useful for this case!
With our Workflows tool (Roboflow Workflows: The open source tool to build vision applications) you can set up a two stage model workflow that can stream on edge with our Inference Pipeline. As you mention, this will be slower than one model, so I’m not sure if it will work for your use case. But this will definitely be the best option for accuracy!
Otherwise, manually creating each keypoints class should work, even though the setup will need to be done manually for each new class name.
Thanks, I will probably just create a separate classification model and use it to change classes on the keypoint data