PoseEstimationForMobile: A C++ repository from iMageChans

This repository currently implemented the CPM and Hourglass model using TensorFlow. Instead of normal convolution, inverted residuals (also known as Mobilenet V2) module has been used inside the model for real-time inference.

Model	FLOPs	PCKh	Inference Time
CPM	0.5G	93.78	~60 FPS FPS on Snapdragon 845 30+ FPS on iPhone x (need more test)
Hourglass	0.5G	91.81

You can modify the architectures of network for training much higher PCKh model.

The respository contains:

Code of training cpm & hourglass model
Android demo source code
IOS demo source code

Below GIF is catch on Mi Mix2s (~60 FPS)

You can download the apk as below to test on your device.

Using Mace (Support GPU)	Using TFlite (Only CPU)
PoseEstimation-Mace.apk	PoseEstimation-TFlite.apk

Issue and PR are welcome when you encount any problem.

Training

Dependencies:

Python3
TensorFlow >= 1.4
Mace

Dataset:

Training dataset available through google driver.

Unzip it will obtain the following file structure

# root @ ubuntu in ~/hdd/ai_challenger
$ tree -L 1 .
.
├── ai_challenger_train.json
├── ai_challenger_valid.json
├── train
└── valid

The traing dataset only contains single person images and it come from the competition of AI Challenger.

22446 training examples
1500 testing examples

I transfer the annotation into COCO format for using the data augument code from tf-pose-estimation respository.

Hyper-parameter

In training procedure, we use cfg file on experiments folder for passing the hyper-parameter.

Below is the content of mv2_cpm.cfg.

[Train]
model: 'mv2_cpm'
checkpoint: False
datapath: '/root/hdd/ai_challenger'
imgpath: '/root/hdd/'
visible_devices: '0, 1, 2'
multiprocessing_num: 8
max_epoch: 1000
lr: '0.001'
batchsize: 5
decay_rate: 0.95
input_width: 192
input_height: 192
n_kpoints: 14
scale: 2
modelpath: '/root/hdd/trained/mv2_cpm/models'
logpath: '/root/hdd/trained/mv2_cpm/log'
num_train_samples: 20000
per_update_tensorboard_step: 500
per_saved_model_step: 2000
pred_image_on_tensorboard: True

The cfg not cover all the parameters of the model, there still have some parameters in the network_mv2_cpm.py.

Train by nvidia-docker

Build the docker by the following command:

cd training/docker
docker build -t single-pose .

docker pull edvardhua/single-pose

Then run the following command to train the model:

nvidia-docker run -it -d \
-v <dataset_path>:/data5 -v <training_code_path>/training:/workspace \
-p 6006:6006 -e LOG_PATH=/root/hdd/trained/mv2_cpm/log \
-e PARAMETERS_FILE=experiments/mv2_cpm.cfg edvardhua/single-pose

Also, it will create the tensorboard on port 6006. Beside, make sure you install the nvidia-docker.

Train by ordinary way

install the dependencies.

cd training
pip3 install -r requirements.txt

Beside, you also need to install cocoapi

Edit the parameters files in experiments folder, it contains almost all the hyper-parameters and other configuration you need to define in training. After that, passing the parameters file to start the training:

cd training
python3 src/train.py experiments/mv2_cpm.cfg

After 12 hour training, the model is almost coverage on 3 Nvidia 1080Ti graphics cards, below is the corresponding plot on tensorboard.

Bechmark (PCKh)

Run the follow command to evaluate the value of your PCKh.

python3 src/benchmark.py --frozen_pb_path=hourglass/model-360000.pb \
--anno_json_path=/root/hdd/ai_challenger/ai_challenger_valid.json \
--img_path=/root/hdd \
--output_node_name=hourglass_out_3

Pretain model

CPM

Hourglass

Android Demo

Thanks to mace framework, now you can using GPU to run this model on android smartphone.

Following command can transfer model into mace format.

cd <your-mace-path>
# You transer hourglass or cpm model by changing `yml` file.
python tools/converter.py convert --config=<PoseEstimationForMobilePath>/release/mace_ymls/cpm.yml

Then follow the instruction of mace documentation to integrate into android.

For how to invoke the model and parsing output, you can check the android source code i provided.

The benchmark of some socs for average inference time are shown as follow.

Model	Snapdragon 845	Snapdragon 660	Hisilicon 960	Exynos 7420
CPM & Hourglass	17 ms	30 ms	42 ms	103 ms

Or you can transfer the model into tflite.

# Convert to frozen pb.
cd training
python3 src/gen_frozen_pb.py \
--checkpoint=<you_training_model_path>/model-xxx --output_graph=<you_output_model_path>/model-xxx.pb \
--size=192 --model=mv2_cpm_2

# If you update tensorflow to 1.9, run following command.
python3 src/gen_tflite_coreml.py \
--frozen_pb=forzen_graph.pb \
--input_node_name='image' \
--output_node_name='Convolutional_Pose_Machine/stage_5_out' \
--output_path='./' \
--type=tflite
 
# Convert to tflite.
# See https://github.com/tensorflow/tensorflow/blob/master/tensorflow/docs_src/mobile/tflite/devguide.md for more information.
bazel-bin/tensorflow/contrib/lite/toco/toco \
--input_file=<you_output_model_path>/model-xxx.pb \
--output_file=<you_output_tflite_model_path>/mv2-cpm.tflite \
--input_format=TENSORFLOW_GRAPHDEF --output_format=TFLITE \
--inference_type=FLOAT \
--input_shape="1,192,192,3" \
--input_array='image' \
--output_array='Convolutional_Pose_Machine/stage_5_out'

Then, place the tflite file in android_demo/app/src/main/assets and modify the parameters in ImageClassifierFloatInception.kt.

......
......
    // parameters need to modify in ImageClassifierFloatInception.kt
    /**
     * Create ImageClassifierFloatInception instance
     *
     * @param imageSizeX Get the image size along the x axis.
     * @param imageSizeY Get the image size along the y axis.
     * @param outputW The output width of model
     * @param outputH The output height of model
     * @param modelPath Get the name of the model file stored in Assets.
     * @param numBytesPerChannel Get the number of bytes that is used to store a single
     * color channel value.
     */
    fun create(
      activity: Activity,
      imageSizeX: Int = 192,
      imageSizeY: Int = 192,
      outputW: Int = 96,
      outputH: Int = 96,
      modelPath: String = "mv2-cpm.tflite",
      numBytesPerChannel: Int = 4
    ): ImageClassifierFloatInception =
      ImageClassifierFloatInception(
          activity,
          imageSizeX,
          imageSizeY,
          outputW,
          outputH,
          modelPath,
          numBytesPerChannel)
......
......

Finally, import the project to Android Studio and run in you smartphone.

iOS Demo

Thanks to tucan, now you can run model on iOS.

First, convert model into CoreML model.

# Convert to frozen pb.
cd training
python3 src/gen_frozen_pb.py \
--checkpoint=<you_training_model_path>/model-xxx --output_graph=<you_output_model_path>/model-xxx.pb \
--size=192 --model=mv2_cpm_2

# Run the following command to get mlmodel
python3 src/gen_tflite_coreml.py \
--frozen_pb=forzen_graph.pb \
--input_node_name='image' \
--output_node_name='Convolutional_Pose_Machine/stage_5_out' \
--output_path='./' \
--type=coreml

Then, follow the instruction on PoseEstimation-CoreML.

Reference

[1] Paper of Convolutional Pose Machines
[2] Paper of Stack Hourglass
[3] Paper of MobileNet V2
[4] Repository PoseEstimation-CoreML
[5] Repository of tf-pose-estimation
[6] Devlope guide of TensorFlow Lite

License

Apache License 2.0

iMageChans/PoseEstimationForMobile