Regression of just a single (1) keypoint per instance - should I use the pose model or attempt my own modification?

[edavidk7]

I have a particular computer vision pipeline. Part of my pipeline is the detection of colored traffic cones in the image. Currently, I just use a regular YoloV8-small trained on the dataset, which gives me the class and the bounding box around the traffic cone. Further in the pipeline, I need to estimate the center of the base of the cone (which lies on the ground) and its respective image coordinates (i. e. where it would have projected in the image if it were visible).

I would like to move the estimation of the cone base center to the YOLO network to make it much more general and immune to changes in camera parameters, as the current method is just a polynomial that fails if the parameters of the camera change even slightly.

I have a method for relatively accurately labeling the data of this center for a given bounding box, meaning I can create a few hundred to low thousands of data samples.

I know there is the possibility of using the pose estimation model, which can take 2D points for regression.
Can anyone with more experience with this model say whether it would be adequate for this task? Or should I attempt to make my own modification for better performance?

Also, is there any possibility of transferring the weights of my already trained model into a blank, randomly initialized pose estimation model?

Thanks a lot!

[glenn-jocher]

For estimating the center of the base of traffic cones, you could indeed leverage the pose estimation capabilities of YOLOv8, as it is designed to detect specific points in an image. Given that you only need to regress a single keypoint per instance, the pose model could be a good fit for your task, especially since you have a method to label the data accurately.

However, if your requirements are very specific and you find that the pose model does not meet your needs, you might consider a custom modification. Since you already have a trained YOLOv8 model, you could explore adding a regression head to directly predict the keypoint coordinates from the detected bounding box.

Regarding weight transfer, yes you can attempt to transfer weights from any model to any model using .load(), i.e.

model = YOLO('yolov8n-pose.pt').load('your/custom/detection/model.pt')

Feel free to experiment with both approaches to determine which yields the best results for your specific use case. Good luck with your project! 🚀