[Exynos/Mali] ExternalTextureConverter issue with hand tracking sample App

[cheongjunyoung]

[Exynos/Mali] ExternalTextureConverter issue with hand tracking sample App

[leetenthu]

Is there a corresponding TF bug tracing this so that we can follow up?

[mcclanahoochie]

Hi,
I was able to get two galaxy s8 phones for testing, one with Adreno GPU and one with Mali GPU. I can reproduce this issue on the Mali chipset, while the Adreno works just fine.

It appears there is some hardware/memory limitations on the galaxy s8 G950FD with Mali chipset.
On that phone, none of the GPU examples with tflite seem to work as-is, but i was able to modify the CPU examples to run inference on the GPU (i.e. add "use_gpu: true" to the TfLiteInferenceCalculatorOptions in the graphs.

You may be able to create a CPU version of the hand graphs (remove all "_GPU" from stream names in the graphs) and run inference on the GPU still (add "use_gpu:true" to inference calculator options), as a workaround, but the performance won't be that good.

Meanwhile we will continue to improve support for a variety of devices.

[sevar83]

Same on Galaxy S9.

[Macoron]

Reproduced on Galaxy S9, though Xiaomi Redmi 4x works fine

[leetenthu]

Same issue on Galaxy S9+ (SM-G965F/DS).

[mgyong]

This bug is specific to TF. Will close in MediaPipe

[chuoling]

We (MediaPipe) will follow up with TensorFlow team and provide an update here.

[hkawii]

Any updates on that issue ?

Same on S8+ and A20

[mgyong]

@hebamohamed @leetenthu @Macoron @sevar83 @cheongjunyoung Fixed bug in v0.6.5 release. Pls check out v0.6.5 and let us know if this release fixed the issue

[VRShard]

Tested on Mali G51 multihandtrackinggpu --3D=true, on my side it is working now.

[leetenthu]

Tested on Galaxy S9+, it is working now, thanks!