JONGGON / Deephumanprediction

Introduction

• A place to 'post' the progress of my master's thesis.

Progress(Related studies necessary for master 's thesis.)

• Where can I get the data?

  • https://accad.osu.edu/research/mocap/mocap_data.htm
  • http://mocapdata.com/

• what is BVH file?

• Reading BVH file?

  • Code

  • PreProcessing the data Using Motionbuilder - Click on this Sentence and Learn

    • Since the skeleton information of 'ACCAD' and MOCAPDATA.COM data is different, each preprocessing is necessary.(ACCAD dataset -> Used in Progress , MOCAPDATA.COM dataset -> used in Master's Thesis)

    • Convert from C3D to BVH using MotionBuilder - It can be difficult because it uses a professional program called 'Motion Builder'.

• Writing BVH file?

  • Code

• Human Motion Prediction basic -Various attempts

  • Encoding + Decoding Structure

    • Motion Prediction Simple - completed

    • Motion Prediction encoding decoding - completed

      • You have to learn one by one from easy data.(Please refer to the code.)

    • Motion Prediction encoding decoding using BidirectionalCell - completed

      • You have to learn one by one from easy data.(Please refer to the code.)

    • Motion Prediction encoding decoding : Joint angle to Cartesian coordinates using open source - Fixing

      • You have to learn one by one from easy data.(Please refer to the code.)

    • Motion Prediction decoding encoding : using training_set and test_set - completed

      • You have to learn one by one from easy data.(Please refer to the code.)

    • Motion Prediction basic Seq2Seq sequencial learning version - completed

      • You have to learn one by one from easy data.(Please refer to the code.)

    • Motion Prediction advanced Seq2Seq sequencial learning version - completed

      • You have to learn one by one from easy data.(Please refer to the code.)

  • Encoding + Decoding Sequence to Sequence Structure

    • Motion Prediction advanced Seq2Seq batch learning version - completed

    • Motion Prediction Seq2Seq - Not Using Decoder Input - completed

Master's Thesis

• Human motion for Master's Thesis

  • 1-1. Motion Prediction(Generation) , Batch Learning , Input with noise , Tied weights , Training method = Test method

  • 1-2. Motion Prediction(Generation) , Batch Learning , Input with noise , Tied weights , Training method != Test method

    '''Please read the notes below'''
    
    * about Network structure
    
        - If we share encoder and decoder weights, we can only use Residual Connection for 'One RNN Layer', 
        because 'ResidualCell' must have the same size of input and output.
    
        - The code for MultiLayer is not actually used, but is written for later use.
    
    * about Dataset
    
        - <Training data> : The training motion data (272) for 4 women, 68 motion data per person.
    
        - <Test data> : The Test motion data for 1 women, 68 motion data per person.
    

  • Human Motion Analysis consists of 3-Projects

    - <Training data> : The training motion data (1224) for 10 men and 8 women, 68 motion data per person.
    
    - <Test data> : The Test motion data (204) for 2 men and 1 women, 68 motion data per person.
    

    • 1. Motion Classification

    • 2. Human Motion Pattern Classification

    • 3. Motion Prediction(Generation) With more data

      Question? Can the network characterize motion data? Is it possible to generalize the Human motion?
      

Development environment

• window 10.1 64 bit and Ubuntu Linux 16.04.2 LTS
• python verison : 3.6.1 , anaconda3 version : (4.4.0)
• pycharm Community Edition 2017.2.2

Dependencies

• mxnet-0.12.1(window) , mxnet-0.12.1(Linux)
• tqdm -> (progress) , graphviz -> ( Visualization )