pzoltowski/ecpose-s-onnx

EdgeCrafter-S pose (ONNX)

EdgeCrafter ECPose-S one-stage multi-person pose model (17 COCO keypoints) as ONNX, for OpenCV dnn and the react-native-opencv demo.

• Source: EdgeCrafter (https://github.com/Intellindust-AI-Lab/EdgeCrafter), via the LibreYOLO checkpoint LibreYOLO/LibreECs-pose, exported to ONNX with named outputs.
• License: Apache-2.0. The weights are © The EdgeCrafter Authors (Intellindust AI Lab); this mirror redistributes them with attribution (see LICENSE and NOTICE). This is not the non-commercial -obj2coco variant.
• Input: 1×3×640×640, float32 RGB.
• Outputs: pred_logits (Q,2) (sigmoid of the last column = person score) and pred_keypoints (Q,34) (x,y normalised to [0,1]). No per-keypoint confidence and no box head; the box is the keypoint extent.
• Verified in OpenCV 5 DNN against the PyTorch reference: keypoint error ≤ 0.0013 px on 4 test images. Decoder: cv::supavision::decodeKeyPointsEcPose.
• File: LibreECs-pose.onnx, 41,137,581 bytes, sha256 45a31e81f8c3ca435789cee193b5b9a43e28f78af65573434ac74d60b1b4e732.