vision_flow 0.0.4

A Flutter plugin for real-time vision tasks, Hands, Face, Pose estimation, and Video Classification. Built on top of MediaPipe, powered by ExecuTorch and TensorFlow Lite.

Changelog

0.0.1

• complete architectural overhaul and renaming from signify_vision to VisionFlow.
• added VisionFlow.loadModel() to select your model file and backend dynamically.
• added VisionFlow.configure() to choose hands, face, pose, and sequenceLength dynamically.
• modular backend architecture: added generic ModelRunner interface supporting both PyTorchRunner and TFLiteRunner.
• feature extraction pipeline: modified MediaPipe Hand/Face integration to output specifically 330 coordinates per frame matching the WLASL notebook model format.
• sequence normalization: Built-in sequence normalization across sequenceLength buffers directly in FrameBuffer, correctly centering landmarks on the nose relative point per sequence.
• Flutter Event Streams: Changed predictions API to return a formatted PredictionResult with classification index.

0.0.2

• updated example/ folder.
• updated Documentation.
• updated LICENSE file.

0.0.2+1

• updated pubspec.yaml.
• updated README.md.
• updated LICENSE file.

0.0.3

• added file picker for loading model files.
• fixed bug with AndroidManifest.xml.

0.0.3+1

• Upgraded TensorFlow Lite and PyTorch dependencies.
• Upgraded file_picker from v8.0.0 to v11.0.2.

0.0.4

• Breaking: Switched PyTorch backend from pytorch_android_lite to ExecuTorch 1.3.1 (org.pytorch:executorch-android:1.3.1).
• PyTorch model format changed from TorchScript (.pt) to ExecuTorch portable format (.pte). Re-export your models using torch.export + to_edge_transform_and_lower.
• Updated PyTorchRunner.kt to use the new org.pytorch.executorch API (Module, EValue, Tensor).
• Updated example app file picker to accept .pte instead of .pt.
• Updated README.md with ExecuTorch model export instructions.