I am trying to train a model to detect key points on a pickleball court. My dataset has about 1135 images with annotations. I tried to use both Roboflow 3.0 Keypoint Detection (Fast) and YOLOv11 Keypoint Detection (Fast) to detect the key points.
However, the model does not seem to be able to identify the key points well. How can i improve this? Pretty new to CV. Appreciate any inputs from the community.
The training data is pretty much the same court as the test data. Which is why i am surprise about the test result. I thought i might be overfitting. But it does not turn out that way in the test set. In my screenshot, this is the test result. You can view it from my workspace.
For production, there is a little difference with the audience on the sidelines. but pretty much the same view of the court.
Hi @Desmond_Wong!
Your annotations look fantastic!! In looking at your test set, it looks like the model trips up when a keypoint is occluded. So I would suggest adding some more training images of that case.
I would add some diversity to your dataset using a grayscale augmentation. This will force the model to learn to key in on other features other than color to make predictions.
