TODO.md

TODO

Overall

• Research research strategies
• Email EIT for download exemption
• Download http://horatio.cs.nyu.edu/mit/tompson/nyu_hand_dataset_v2.zip
• Clone https://github.com/jsupancic/deep_hand_pose and run some code
• Reimplement Deep Hand Pose in TensorFlow
  • Build the graph structure
    • Read the graph protobuffer format, understand the graph layers & vector input/output shapes, document
    • Translate graph into Keras graph
    • Run and debug Keras graph
  • Convert the NYU dataset to something TensorFlow can read
    • Look at converting .png to TensorFlow vectors (TFRecord files?)
    • Figure out how to add images to .npz file individually
    • Read labels and understand/document their format
    • Write script to convert the NYU dataset into a single HDF5 archive
• Make a short slideshow
  • Overview/background of project
  • Problem statement (inputs, outputs, process)
  • Goals (mobile, non-depth)
• Write a paper draft
  • Synthesize information from slideshow
  • Collect sources
  • Document research process
  • Work on a journal / log
  • Report on preliminary results (if any)
• Experiment with improvements
  • Augment data with scales and translations
    • Refactor for performance
  • Add label depths
  • Add batch normalization
  • Use autoencoder instead of PCA
  • Try more PCA components
  • Use striding instead of max pooling
  • Change number of filters
  • Change batch size
  • Change depth of network
  • Change neurons in fully connected layers
• Work on mobile port
• Document and publish
  • Create a GitHub page
  • Generate results graphs
  • Add model comparisons
  • Start results section
  • Add installation documentation
  • Add 'why' to code documentation

This week

• Try more PCA components
• Change batch size
• Start results section
• Add model comparisons