TALEZ Dog Motion ยท 2026-10-20 Google MediaPipe Tasks Vision / EfficientDet Lite0: existing site-pinned package and model. https://developers.google.com/edge/mediapipe/solutions/vision/object_detector Animal pose: OpenMMLab RTMPose-m, AP-10K, Apache-2.0. Original: https://download.openmmlab.com/mmpose/v1/projects/rtmposev1/onnx_sdk/rtmpose-m_simcc-ap10k_pt-aic-coco_210e-256x256-7a041aa1_20230206.zip Model: end2end.onnx (54,478,120 bytes) SHA256: 1cfd1c86e0d9e5d5f95178bcd95ee9a4e8386a624cd3c57519f27ff58cac7f28 Runtime URL is a revision-pinned Hugging Face mirror; byte-equality with official ONNX verified 2026-10-20. No model conversion or retraining. https://huggingface.co/hr16/UnJIT-DWPose/tree/9a04e1ebb32c2cb7526f89933b774f367242e9f5 Original license: /dog-motion/LICENSE-MMPose.txt AP-10K: Hang Yu, Yufei Xu, Jing Zhang, Wei Zhao, Ziyu Guan, Dacheng Tao. AP-10K: A Benchmark for Animal Pose Estimation in the Wild. NeurIPS 2021 Datasets and Benchmarks. https://github.com/AlexTheBad/AP-10K (CC BY 4.0) Dog label correction, 2026-10-22: the official AP-10K README (Key Questions 4) defines generic hip/knee labels as knee/ankle regions for digitigrade animals. We now display and export knee/hock regions for the dog's hindlimbs. These are estimated image regions, not measured anatomical joint centres. No hip joint or forelimb wrist centre is separately estimated. No weight conversion or retraining. ONNX Runtime Web 1.21.0, Microsoft Corporation, MIT. https://github.com/microsoft/onnxruntime/blob/v1.21.0/LICENSE Version-pinned browser runtime and matching WASM files are fetched from jsDelivr. TALEZ modifications: independent browser preprocessing, confidence/temporal rejection, 2D overlays, local video/camera controls and descriptive results. Not medically validated. No 3D, force, pain, or muscle-activation measurements. No video or keypoints sent to external inference APIs. 2026-10-22 accuracy hold: the model's peak scores and candidate counts do not validate anatomical locations or occlusion. Skeleton lines and numeric angle reports are withheld. Candidate dots are hidden by default and available only for explicit review; exported rows are marked unvalidated. The model has not been retrained or validated against anatomical ground truth. 2026-10-22: File candidates are shown only on their paused analysed timestamps; playing/seeking clears the overlay. External research marker comparisons are documented in the source validation/dog-motion folder. They are not anatomical validation; weights and the runtime detector are unchanged.