Pose Model Compatible with Oak-D Pro

AidanKwok

Are there any Yolo pose models that can run on the Oak-D pro? I tried using the blob converter to convert various Yolo pose .pt to .blob but ran into formatting issues. Ideally, I'd be able to extract the pose key points and use the bounding box for depth. I saw the human-pose-estimation-0001 model with the DepthAI SDK but read that the inference times were slow. I'm also not sure how accurate it is at distance. I need to detect pose at 10 meters.

KlemenSkrlj

Hi @AidanKwok ,
Yes, we support multiple pose models natively on OakD Pro. There are some that are already exported in our Model Zoo and ready to use. But we also support conversion process from multitude of YOLO Ultralytics models. You can check the HubAI conversion process here and the current list of supported models here.
Related to the speed: The YOLOv8 nano instance pose for example is benchmarked on RVC2 at _33FPS so its real-time.
As for the distance: This depends on the environment and how well the model works in your deployment. One way to improve the detection of small objetcs at the cost of some inference speed though is to try out the Tiling approach where you are esentially split the image into tiles, predict on each tile separately and then merge the predictions back to singe one.

But note that all of this is supported with DepthAI v3, using .superblob versions of the models. We would anyway recommend that you switch to using DepthAI v3 since this is now actively being maintained and developed for. You can check the migration guide for some help.
Best,
Klemen