Potential Topics

• detect computer keyboard key hits from video; this is probably only feasible for a reduced set of allowed keys
  • possible approaches:
    • using space time volumes (Filip Ilic PhD) of the hand
    • directly using image segmentation (detecting shape of the hands) and learning individual key presses from these (stationary) images with NN
  • data source:
    • self made
  • potential issues:
    • distinguish pressing/touching keys
• detecting different routines from platfom/cliff diving videos
  • possible approaches:
    • using space time volumes (Filip Ilic PhD)
  • data source:
    • videos from red bull cliff diving world series competitions
  • potential issues:
    • ?
• cloud forecast based on satellite image sequence or rain forecast based on weather radar image sequence
  • possible approach:
    • R(C)NN
  • data source:
    • one of the datasets from kaggle
    • ZAMG, or equivalent …
• If we want to keep the effort low, we could do something similar to the examples in Chapters 11-14 from this book: Deep Learning Cookbook, Douwe Osinga

Space-Time Volume Gesture Classification

Image Preprocessing

• morphological operations if necessary

Feature Extraction

Maybe do a principal component analysis on the potential features to find most descriptive features.

1. extract contours
   • perform pointPolygonTest to detect overlapping/occluding volumes
   • contour tracking? from lecture slides: Kass, Witkin, Terzopoulos, 1st ICCV, London, 1987
2. fit shape
   • fitEllipse
   • fitLine
   • minAreaRect
3. compute scalar feature values
   • for 2-3 largest shapes:
     • absolute orientation angle of two biggest objects
     • relative orientation angle
     • relative distance
     • distance to center of bounding box
     • relative size w.r.t. bounding box
     • elongation/compactness
   • number of shapes in bounding box
     • with relative size/area w.r.t. bounding box greater than some threshold

Feature normalization/centering: could become interesting with the absolute angle as 0=pi=2pi -> maybe rough discretization makes sense or (co)sine

Feature/Token Selection - Outlier Suppression

• minimum relative area of fitted shape w.r.t. boundingbox
• maybe ransac

Feature Filtering

• matchShapes for STV contours over time:
  • if there are discontinuities in the matching value -> interpolate
  • or: use it for RANSAC
• median filtering for scalar feature values

ML

• CNN with 1D convolution layers and pooling layers
• fully connected NN on entire feature-over-time sequences
• RNN ?
• Cost function: cross entropy

Code structure

• Data preprocessing:
  • classifying-tool.py
  • visualization-tool.py
• Feature extraction:
  • batch_extract_features.py
• Training:
  • svm_train_test.py
• Evaluation/visualisation:
  • test.py: run feature extraction on individual video
  • evaluate.py: visualize feature extraction & classification result of individual video
  • box_plot_accuracy.py