Computer vision task
Keypoint and pose datasets and models
Landmark and pose datasets with per-instance skeletons.
Keypoint detection locates specific landmarks on an object - the joints of a body for pose estimation, facial features, or the corners of a part. Each keypoint is a labelled point, and points that belong to the same object are grouped so a model can reconstruct a full skeleton or pose.
Pictograph stores keypoints in the COCO keypoints convention and groups the joints of one object with an instance id, with the skeleton template defined once per class. That is exactly the supervision a top-down, query-based model needs to learn multi-instance pose, so datasets export straight to COCO-Pose and are ready to train.
Pictograph trains keypoint models with RF-DETR and packages them as ONNX. The dataset(s) below carry keypoint annotations you can fork and extend.
Datasets 1
Keypoint detection FAQ
What is a keypoint or pose dataset?
A keypoint dataset labels specific landmarks on each object - body joints, facial features, or part corners - as individual points, with the points of one object grouped together. Trained on it, a model predicts those landmarks on new images, which is the basis of pose estimation and fine-grained localisation.
What format does Pictograph use for keypoints?
Pictograph follows the COCO keypoints convention: each joint is a class, an instance id groups the joints of one object, and the skeleton (which joints connect) is defined once per class rather than repeated on every annotation. Datasets export directly to COCO-Pose.
How do I train a keypoint model on Pictograph?
Fork or upload a dataset with keypoint annotations grouped by instance id, then start a training run with the RF-DETR keypoint pipeline. Pictograph runs the export, GPU training, and ONNX packaging, and serves the model through the API or as a downloadable file.
Start building today
$5/mo free compute. No payment method required.