Hi. I love yolo since 5 years, 
ultralytics for efforts.
-I’ve prepared a test dataset of 250 images for a medical project where keypoints unrelated to the human body are important.
-I annotated it with Coco annotator in xyv format for YOLO Pose v8.
-I converted the JSON file to YOLO format.
-I trained it from scratch using yolov8n-pose.yaml.
-When I infer with the resulting PT model using results[0].plot(), it also draws lines connecting the keypoints.
My question is:
-
When the model detects keypoints, does it only use the annotation data for that specific keypoint, or does the relative position of other keypoints affect its placement? In other words, is there any effective skeletal system or algorithm that considers the distance between keypoints? Because I only need accurate detection of the correct keypoint in the image, not influenced by a skeletal system or other keypoints. The poor detection of keypoints is likely due to having a limited amount of annotated images, not the skeletal system, right?
-
Since the project will use a 1920x1200 camera, I trained it at a resolution of 1920. Do you think YOLOv8n is suitable for higher resolutions? Or any other letter than “n”
-
My camera is monochrome. I don’t want to use grayscale because it complicates things, but would switching the model from (h,w,3) to (h,w,1) significantly speed it up? Is it worth the effort?"
Best regards:heart: