nrl-ai/anylearning-rfdetr-locust-keypoints

AnyLearning RF-DETR Desert Locust Keypoints

An RF-DETR Keypoint Preview checkpoint trained and validated end to end in AnyLearning OSS. It detects one Desert Locust and predicts 35 named anatomical landmarks.

Files

• model_best.pth — native RF-DETR checkpoint
• exported_model.onnx — checked ONNX graph with a fixed 1 × 3 × 576 × 576 input and box, class and 35-keypoint outputs
• config.yml — landmark order, thresholds and training configuration
• labels.json — AnyLearning project schema, including 26 skeleton edges
• metrics.json — machine-readable validation and held-out-image measurements

Training data

The public Desert Locust archives contain 630 training and 70 validation images from DeepPoseKit-Data. Both archives embed their Apache-2.0 attribution and the Graving et al. (2019) citation. Images are 160 × 160 pixels with one annotated locust per image.

Training continued the official RF-DETR Keypoint Preview checkpoint for 5 epochs and then for 10 more epochs with batch size 2, gradient accumulation 4, learning rate 5e-5, 576-pixel model input, bfloat16 mixed precision and horizontal flipping. The left/right landmark pairing used by the flip is recorded in config.yml.

Evaluation

Final epoch metrics on the 70-image validation split:

Metric	Value
Keypoint mAP (50:95)	0.846
Keypoint mAP@50	1.000
Keypoint mAP (EMA)	0.849
Box mAP (50:95)	0.975
Box mAP@50	1.000

An additional application-level inference check used one validation image that was excluded from training. AnyLearning reloaded the registered native model, returned one box and all 35 landmarks, and rendered the prediction. Against the stored ground truth, mean error was 2.45 px, median error 1.70 px, PCK@5 px 91.4%, PCK@10 px 97.1%, and maximum error 11.71 px on the 160 × 160 image.

These figures describe this split only. The dataset has a single individual, fixed image size and controlled grayscale imagery. Evaluate on your own camera, species, pose range and occlusion patterns before relying on it.

Checksums

File	SHA-256
model_best.pth	7ccfdfeb72ea59aa601678f8699ffa7b806009148b9edd7534e0595850461b58
exported_model.onnx	70f5ed0fe988ed2e7480d9b59aa5776224cc6564fcf93f898c35fe06ba04983a

Licence and citation

The RF-DETR code and released Keypoint Preview weights are Apache-2.0, and the DeepPoseKit-Data repository publishes its example datasets under Apache-2.0. Preserve the supplied attribution and cite:

Graving, J. M. et al. (2019). DeepPoseKit, a software toolkit for fast and robust animal pose estimation using deep learning. eLife, 8, e47994. https://doi.org/10.7554/eLife.47994