Hourglass model

[edvardHua]

Try to use hourglass model in here.

• input_image_shape: [1, 192, 192, 3]
• output_shape: [1, 48, 48, 14]
• input_node_name: image
• output_node_name: hourglass_out_3

I think it will speed up 3x than old model and the precision is acceptable.

Tomorrow I will update my repo.

Look forward to your replay ( ? FPS on ipx )

[tucan9389]

@edvardHua AWESOME! I'll try that model 👍 Do you have a python script that converts to Core ML model? I have some issue in step converting to Core ML with my own model now 😓

[edvardHua]

Yes, I will let you know when I update repo tomorrow.

[tucan9389]

@edvardHua I capture video comparing two models. but It's little weird for accuracy.

model	cpm model	hourglass model
inference time	57(45~70)mm	33(30~35)mm
accuracy	more accurate(0.9)	less accurate(<0.5)
DEMO video

[edvardHua]

Indeed, the old one precision is higher than the new one.

PCKh of cpm model is 94.18.
PCKh of hourglass model is 91.

But hourglass do not have too wide difference on android device.

You can test android demo here

[edvardHua]

@tucan9389 Here is the script I used to convert mlmodel.

[tucan9389]

Thanks a lot

[BroderickHigby]

Hey Tucan, Would you mind uploading your mv2_cpm_model_98000 mlmodel? I'm interested in seeing its model accuracy compared to the previous accuracy.

[tucan9389]

@BroderickHigby Unfortunately I can’t open my models. If you want compare between cpm and hourglass model, you can download from here.

[tucan9389]

Close the issue now. You can reopen when you want. @edvardHua