v1.6.0

• Upgrade to ONNX Runtime v1.15.1. The SDK now supports ONNX models with opset version <=19.

What's Changed

• Upgrade ORT to v1.15 by @astahlman in #6

Full Changelog: 1.5.0...1.6.0

1.5.0

• Add support for landscape. This should now work correctly for either portrait or landscape images, or in fact images on any aspect ratio.
• The SDK now includes a copy of a tensorflow object detection model, to improve accuracy for images of different aspect ratios. The model is 7.6mb so should have minimal impact on deployment options.

1.4.0

Add Onnx runtime support and allow loading models from assets/ directory.

What's Changed

• Drop ncnn in favor of ORT by @astahlman in #4

Full Changelog: 1.3.0...1.4.0

1.3.0

Add client-side support for hardware-accelerated models. Also, support tracking the pose without sending the keypoints to Guru for analysis.

What's Changed

• Support ncnn inference and bump to 1.3.0 by @astahlman in #3

New Contributors

• @astahlman made their first contribution in #3

Full Changelog: 1.2.2...1.3.0

1.2.2

• Fixed bug where analysis was sending landmarks in wrong case, resulting in bad analysis results.

1.2.1

• Fixed bug where a partially downloaded model (perhaps because of network disconnect) would then be used as the final model and inference would fail. This will now only use the model if it's fully downloaded, and retry if not.

1.2.0

• Add method overload to allow use of Bitmap for inference.

1.1.0

• Added in pushing of inference to server to enable rep counting and feedback. No video is sent to the server, only spatial coordinates.

1.0.1

• Remove Jitpack dependencies, since it did not seem a reliable distribution method.

1.0.0

Initial release of the SDK. For now this just includes support for live pose estimation.