mediapipe

mediapipe copied to clipboard

I am running the MediaPipe hand tracking pipeline on a resource-constrained edge device with limited CPU capabilities. However, the single-frame inference latency exceeds 1 second, which is impractical for real-time use. To optimize performance, I aim to apply quantization to the models.

1. Obtaining Original Model Files The MediaPipe documentation mentions the need for original model files (e.g., TensorFlow SavedModel or frozen graphs) for quantization. However, the current model repository only provides pre-built TFLite files like hand_landmark_full.tflite and palm_detection_full.tflite. Question: Where can I access the original non-quantized model files (*.pb or SavedModel formats) for hand_landmark and palm_detection?

2. Custom Model Support for Hand Landmark Detection The documentation states that custom models are not officially supported for hand landmark detection. Question: Does this restriction also apply to incremental fine-tuning (e.g., using transfer learning on the existing hand landmark model)? Are there any workarounds to integrate a custom-trained hand landmark model into the MediaPipe pipeline, even if unofficial?

Thank you!!