Mediapipe: Running real-time hand tracking model in EdgeTPU

[Jaspereclipse]

Hi there,

Amazing work on realtime hand-tracking!

Do you have a future plan to provide quantized tflite models to run on EdgeTPU as well? I really want to try out the model on an edge device like raspberry pi with Coral TPU.

Thanks!

[fanzhanggoogle]

Hi, thank you for trying out and good question. We did tried to quantize the models (detector and landmark model), though that part is easy, it turned out the deconv in the detector isn't supported on Coral TPU. We plan to retrain the model without the unsupported op in the near future. But the exact timeline isn't determined yet.

[bhack]

Is It an upsampling or a transposed conv ops (deconv) issue? See also https://github.com/google-coral/edgetpu/issues/31

[bhack]

/cc @warmthless

[bhack]

@fanzhanggoogle:
https://arxiv.org/pdf/1907.01773.pdf
https://github.com/warmthless/split-deconvolution

[bhack]

@fanzhanggoogle Have you tried to give a pass with the new compiler? Check https://github.com/google-coral/edgetpu/issues/45

[bhack]

Check the new list https://coral.ai/docs/edgetpu/models-intro/#supported-operations

[zerometal]

any news on this?

[Jaspereclipse]

Thanks for responding and I am glad to hear you guys have plans to do it. I will keep an eye on the updates.

Thanks again for the great efforts!

[rac0316]

any updates on this!?

[lisbravo]

Still no news on the quantized hand models? As somebody who as invested a great deal of time working on a open source project based on the hand models I will really appreciate having those to resume my efforts.

[garyscetbon]

I'd love a tflite 8-bit quantization of the hand tracking model as well. Any update on this?

[dreamibor]

@fanzhanggoogle Hi, are you planning to share the source code for those trained models? So we can use the source code to quantize those models by ourselves? Thanks.