Experiments

[tucan9389]

Table

• Experiment02
• Experiment03
• Experiment04

[tucan9389]

Experiment02

Dataset Config

• ai_challenger-gpu.cfg

Training Config

• cpm mv2: experiment02-cpm-gpu.cfg
• hourglass mv2: experiment02-hourglass-gpu.cfg
• simplepose resnet18: experiment02-sp-resnet18-gpu.cfg

Result

Summary

Following PCKh@.5s are evaluated with AI Challenger's validation set.

Name	Backbone	PCKh@.5	tflite size	input shape	output shape
CPM	MobileNetV2	92.91%	4.1MB	[1,192,192,3]	[1,24,24,14]
Hourglass	MobileNetV2	72.58%	1.9MB	[1,192,192,3]	[1,48,48,14]
SimplePose	ResNet18	97.22%	61.5MB	[1,192,192,3]	[1,48,48,14]

Tensorboard

Preparing...

[tucan9389]

Experiment03

models from https://github.com/edvardHua/PoseEstimationForMobile/tree/master/release

Result

Summary

Following PCKh@.5s are evaluated with AI Challenger's validation set.

Name	Backbone	PCKh@.5	tflite size	input shape	output shape
CPM	MobileNetV2	96.29%	2.4MB	[1,192,192,3]	[1,96.96,14]
Hourglass	MobileNetV2	94.93%	1.7MB	[1,192,192,3]	[1,48,48,14]

[aitikgupta]

The values of PCKh@0.5 and tflite size seem to be empty for Experiment03. Are you planning to update it?

[tucan9389]

@aitikgupta

The values of PCKh@0.5 and tflite size seem to be empty for Experiment03. Are you planning to update it?

Sure. I update it just now.

[aitikgupta]

For Experiment02, where could I find the configuration and tflite files?

[tucan9389]

For Experiment02, where could I find the configuration and tflite files?

Here are the tflite files. cfg files are placed on config/training folder.

https://github.com/tucan9389/tf2-mobile-pose-estimation/releases/tag/experiment02

[aitikgupta]

Thanks, could you also release the corresponding frozen .pb files too?
I wanted to compare them with my trained model.

[tucan9389]

Thanks, could you also release the corresponding frozen .pb files too?
I wanted to compare them with my trained model.

I just have saved_model now. If you open new issue for this, I’ll share soon.
Thank you.

[tucan9389]

Experiment04

• various stage number of CPM models
• input shape: [1,192,192,3]
• output shape: [1,24,24,14]

Hyperparameters

• number_of_epoch: 200 (about 140.2k step)
  → need to experiment with bigger epoch in the production model

Results

Stage num	PCKh@.5	tflite size
1	60.16%	2.8MB
2	60.24%	3.3MB
3	62.38%	3.7MB
4	93.27%	4.2MB
5	92.82%	4.7MB

Experiment05

• simple pose with various MobileNetV2

#	Channels	PCKh@.5	tflite size
1	[256, 256]	77.58%	9.3MB

Comments

The first PCKh@.5 doesn't satisfy. I'll reconstruct and experiment again.