DYLEM-GRID: Dynamic Leap Motion Gesture Recognition Indexed Dataset

DYLEM-GRID is a dynamic hand-gesture recognition dataset of 400 recordings captured with a Leap Motion controller, covering 4 gesture classes (air_quotes, finger_wagging, waving, zoom). Each recording is a time series of skeletal hand features (palm position, velocity, orientation, width and the full joint tree of the five fingers) stored as one CSV per recording.

Structure

DYLEM-GRID_Raw/          # unprocessed recordings
DYLEM-GRID_Cleaned/      # filtered/aligned version used for the published results
  train/<class>/<id>.csv # 80 recordings per class
  test/<class>/<id>.csv  # 20 recordings per class
DYLEM-GRID_Statistic.csv # per-recording statistics

• Split: 80/20 per class (320 train / 80 test), stratified by gesture
• Features per frame: palm position/velocity/normal/direction/orientation/width + metacarpal..distal joints (position, width, rotation quaternion) for thumb, index, middle, ring, pinky — see the CSV header for the exact column list
• Format: plain CSV (header + one frame per row)

Benchmark

From "A Comparative Analysis of BiLSTM and Transformer Architectures for Dynamic Hand Gesture Recognition" (code):

Model	Accuracy (5-fold CV)
BiLSTM + Attention	97.25% ± 0.94%
Transformer encoder	94.75% ± 0.94%

Attention maps show the BiLSTM learns a temporal filter that focuses on the active gesture phase and ignores idle motion.

Citation

@misc{dylem2024,
  title={DYLEM-GRID---Dynamic Leap Motion Gesture Recognition Indexed Dataset},
  author={Sorce, M. and Lopez, M. A. and Trovato, G. and Cilia, N. D.},
  year={2024},
  publisher={HuggingFace},
  url={https://huggingface.co/datasets/LookUpMark/DYLEM-GRID}
}

License

MIT