Human Pose Classification

Introduction

A project of Human Pose Classification with LSTM.

The final homework of the scoure Artificial Intelligence Comprehensive Experiment in ISE, SYSU.

For more details, please read show.pptx(Simple Chinese).

Teacher: Liu Mengyuan

Group Member: Deng Yushan, Duan Yuxiao, Ding Yuhao, Deng Zhuotao, Yu Zhanpeng

2021/07/25

Human pose sequence

If you want to know more about human pose sequence and see what it like, you cound konw more detial in show.py which is written by nkLiu. You would get the image as before.

LSTM

Synthesizing the original data form and task characteristics, we choose LSTM, a variant of the RNN model dedicated to sequential signal processing in mainstream algorithms. In this task that is highly related to temporal features, LSTM can deal with long-term dependence issues by avoiding long-term memory and avoiding gradient disappearance or explosion problems. It should be more excellent in the classification of time-series sequence signals.

Performance

Speed

GPU	Trainining	GPU Memory
GeForce RTX 2060 Laptop	45 iter/min	4 GB

The training speed is sensitive to your gpu status.

It could be much slower if you do not have GPU and only run with CPU.

Install dependencies

• install TensorFlow, refer to official website. TensorFlow with GPU is faster but not necessery. Code is not GPU only.

  • pip

    pip install tensorflow
    

• install other dependencies:

 pip install numpy 

Load Data

Since the sequences in the data set have different frame numbers, during the process of data loading, this project performs frame-filling operations on the data to increase the frame numbers of all sequences to 350fps. For the method of frame supplementation, this project uses the mean value method, which means that the gap between any two frames is linearly supplemented.

Demo

Download pretrained model from BaiduDisk with password：wtcy

See demo.py for more detail.

You would get the result as below:

Minibatch Loss= 0.15105, 
Testing Accuracy: 0.74358976
Class:000  Test Accuracy: 0.7307692
Class:001  Test Accuracy: 0.7692308
Class:002  Test Accuracy: 0.88461536
Class:003  Test Accuracy: 0.8
Class:004  Test Accuracy: 0.11111111

Train

If you plan to train your own model, you can follow the steps below.

Prepare data

1. Download the training, test data from BaiduDisk with password：stq4

2. Unzip it to the right address

If you are going to use your own data set, please pay attention to the data set format.

Begin training

See train.py for more detail.

Some Key arguments:

• --path the model path
• --learning_rate
• --training_iters
• --batch_size
• --Seq_Len Sequence length
• --n_hidden Number of hidden layers
• --n_classes Number of classes