layout | target | title | parent | nav_order |
---|---|---|---|---|
forward |
Models and Model Cards |
MediaPipe Legacy Solutions |
30 |
{: .no_toc }
- TOC {:toc}
Attention: Thank you for your interest in MediaPipe Solutions. We have ended support for these MediaPipe Legacy Solutions as of March 1, 2023. All other MediaPipe Legacy Solutions will be upgraded to a new MediaPipe Solution. The code repository and prebuilt binaries for all MediaPipe Legacy Solutions will continue to be provided on an as-is basis. We encourage you to check out the new MediaPipe Solutions at: https://developers.google.com/mediapipe/solutions
- Short-range model (best for faces within 2 meters from the camera): TFLite model, TFLite model quantized for EdgeTPU/Coral, Model card
- Full-range model (dense, best for faces within 5 meters from the camera): TFLite model, Model card
- Full-range model (sparse, best for faces within 5 meters from the camera): TFLite model, Model card
Full-range dense and sparse models have the same quality in terms of F-score however differ in underlying metrics. The dense model is slightly better in Recall whereas the sparse model outperforms the dense one in Precision. Speed-wise sparse model is ~30% faster when executing on CPU via XNNPACK whereas on GPU the models demonstrate comparable latencies. Depending on your application, you may prefer one over the other.
- Face landmark model: TFLite model, TF.js model
- Face landmark model w/ attention (aka Attention Mesh): TFLite model
- Model card, Model card (w/ attention)
- Iris landmark model: TFLite model
- Model card
- Palm detection model: TFLite model (lite), TFLite model (full), TF.js model
- Hand landmark model: TFLite model (lite), TFLite model (full), TF.js model
- Model card
- Pose detection model: TFLite model
- Pose landmark model: TFLite model (lite), TFLite model (full), TFLite model (heavy)
- Model card
- Hand recrop model: TFLite model