Keypoint Detection Guide

Keypoint detection finds named landmarks and keeps the points belonging to one object together. The same workflow can describe joints on an animal, corners on a manufactured part, or reference points in a medical image. AnyLearning does not assume a human skeleton: every project defines its own landmark names and connections.

This walkthrough uses the Apache-2.0 DeepPoseKit Desert Locust dataset (opens in a new tab), mirrored as import-ready AnyLearning archives (opens in a new tab). Each 160 × 160 image contains one locust with 35 named landmarks and 26 skeleton edges. The screenshots use the 630-image training and 70-image validation split published by MMPose. Cite Graving et al. (2019) (opens in a new tab) if you use the dataset in research.

Step 1: Create the project and schema

1. Click Create project.
2. Pick Keypoint Detection.
3. Name the project and describe what one instance represents.
4. Add the object class, its ordered keypoint names, and the skeleton edges.

The order is part of the model contract. Landmark 1 must mean the same thing in every image and split; renaming or reordering the schema after labelling makes the annotations inconsistent.

Step 2: Import annotated data

AnyLearning reads COCO keypoint annotations. A category can contain:

• keypoints: the ordered landmark names;
• skeleton: one-based pairs describing which points connect;
• per-object keypoints: [x, y, visibility] triplets; and
• a bounding box identifying the instance.

Put the images and COCO JSON in a zip, then upload the training and validation archives to their matching splits. Keep validation images out of training; a metric measured on images the optimizer has already seen is not a validation metric.

Download train.zip (opens in a new tab) and valid.zip (opens in a new tab), then upload them to their matching subsets. The archives use the documented MMPose 90/10 split and embed their Apache-2.0 attribution.

Step 3: Review or label landmarks

Before training, open images from both splits and check the annotations:

1. Click Start labelling.
2. Select the keypoint tool and choose a landmark name.
3. Assign the same instance number to points belonging to the same object.
4. Mark a point Occluded when its location is known but hidden. Leave a genuinely absent point out instead of inventing a coordinate.
5. Save, reopen the image, and verify the names, instance IDs and visibility.

Use Display to show landmark names, instance IDs, occlusion status and the skeleton. These options change only the canvas; they do not rewrite the saved annotation.

Step 4: Train RF-DETR keypoints

Open Training, click Start training, and choose RF-DETR-Keypoint-Preview.

• Start with the default learning rate, 0.0001.
• Use batch size 2 on a 16 GB GPU. Increase it only after checking memory use.
• Leave image size at Model default for a quality run. A smaller image can make a packaging smoke test faster, but it also removes landmark detail.
• Use enough epochs for held-out keypoint mAP to move. A run that detects the box but reports zero keypoint mAP has not learned the pose yet.
• Compare Validation Keypoint mAP, Keypoint mAP@50, box mAP and loss. No single number explains every failure mode.

The short run in these screenshots is deliberately a pipeline check. It proves that training, export and model registration complete; it is not an accuracy benchmark. For a model you intend to use, train longer and stop only after the held-out landmark metrics and visual predictions are acceptable.

For a reproducible reference, download the public AnyLearning RF-DETR Desert Locust model (opens in a new tab). Its model card includes the native and ONNX checkpoints, exact schema and training configuration, 70-image validation metrics, checksums and a held-out prediction.

Step 5: Inspect and export

1. Open Models and click Try.
2. Choose Use a test image, or upload a genuinely unseen image.
3. Inspect the rendered landmarks and the structured result together.
4. Check for swapped left/right limbs, points attached to the wrong instance, and landmarks that collapse onto a box edge.
5. Download Raw Model for the RF-DETR checkpoint or ONNX Model for the portable graph and its configuration.

Aggregate mAP can hide systematic mistakes, so visual review is part of model evaluation rather than a presentation step. Keep the exported config.yml beside the model: it records the landmark order the output tensor uses.