Goal

The objective is to write a script to decide on the score of a judo throw. Some of the milestones of this project include 1. collecting a large video dataset of games labeled with the referee decisions, 2. finding cool features and 3. training a deep neural network on the data.

Here is an example of a feature that could be used to train the model.

TODO

Joris

• Refactor web scraper

• Fix web scraper

  • Use "judoka" page rather than "ranking"
  • Identify ippon by Hansoku-Make rather than throw

• Characteristics

  • Judoka
    • Family name
    • Given name
    • Country
    • Number of ippons
    • Number of waza
    • Weight category
    • Age
      ...etc
  • Fight
    • is_ippon
    • number waza
    • correct fights with 2 ippons
  • Start organising with Pandas, move to databases later.
  • Multi-threading ?

• See if metadata can be used to store athlete name (mkv)

Yann

• Judo background
• Attention is all you need
• DVC

Ideas

• Features

  • Impact velocity
  • Facing the ground? (Head facing)

• Sequence of images to determine score with number of frames laid_down vs not laid_down

1. Videos

Data management

• Remove folder organization per athlete (all videos in one folder)
• Video title
  • Keep original title for now
  • Use hash (ID) to identify video
  • Parse (=extract info from video title)

2. Image segmentation (CNN)

Ideas

Model Zoo

• Unet* (human shape segmentation)
• Bodypix* TensorFlow.js (human segmentation + body part identification)
• DeepSkeleton (multiple CNN kernels for medial axis detection)
• Holistically-Nested Edge Detection (powerful edge detector ~oriented gradients)

*: pre-trained models available!

Algorithm pipeline

• 1. Segmentation box
• 2. Segmentation silhouette (edge detection)

• keras_segmentation

• 3. Skeletonization

• Apply "grassfire transform" as feature extraction algorithm?

• DeepSkeleton (the source is closed :/)

• 4. Color cue

• "Attention" algorithm

Techniques

• Transfer learning (with fine tuning on our task)
• GPU

3. is_laid_down?

• Difference in Y's between feet and hands.

4. Score detection

I. Score vs not score

II. Waza-ari vs ippon