I’m trying to create a pose model using the ultralytics GUI/platform. I have a little over 50 images. I created just 1 class, “person”, which each point set is assigned to. When I attempt to train the model it falls flat. 0% mAP50-95(P) for the duration of the model training. I run the parent model YOLO26x-pose with an image size of 640. 100 Epochs. Are there any step by step documentation or videos online showing how to train a pose model? Does anybody have any idea what I’m doing wrong? Image attached showing some of the poses and the overlay of points. (I’ve seen the “How to Train YOLO Segmentation Models with Ultralytics Platform” on youtube. Excellent video. Is there something lie that?)
The most likely issue is the dataset rather than the training settings. Each person must have one complete skeleton annotation, with the same 17 keypoints in the same COCO order on every image. Mark hidden points as occluded rather than placing them arbitrarily, and confirm that all validation images contain valid annotations. The Platform pose annotation guide shows the correct Person-template workflow.
With only 56 images and apparently 7 validation images, metrics will also be very unstable. I’d start with YOLO26n-pose or YOLO26s-pose rather than YOLO26x-pose, then add substantially more varied images once the pipeline is verified.
A useful diagnostic is to compare Box and Pose metrics: if Box mAP improves while Pose mAP stays at zero, check keypoint order, visibility, and skeleton consistency. If both stay at zero, check the bounding annotations and train/validation split. You can also clone and train COCO8-Pose on Platform as a quick control test. The Tiger-Pose guide and training video provide a complete custom-pose example.
You have very few images. You would need a lot more. Deep learning requires a lot of images. 1000+
But why do you want to train a pose model? Is it for learning? Because YOLO26 pose can already detect those keypoints. Yout don’t need to train a new model for that.