Automatic DTW-Based Grading for Golf Swing in Sports
how to run
# follow VideoPose3d PoseFormer to download some model weights
# install requirements
pip install -r requirements.txt
# set your path
vim project_config.py
# for GUI
python -m data.gui.gui
# for 2d points visualization
# example (use PE4.mp4)
python -m data.tool.draw_chart data/3d_point_vis/PE4.mp4.npy PE4.mp4
# for analysis see the script folder('data/shell script')
DHU-Golf and data
video -> data/video_raw
grades -> data/grades
hit events label -> data/tag
posture data -> data/3d_point* data/2d_point
model weights
# CBAM-SwingNet
data/golfdb/models/split_4_flip+affine_7700.pth.tar
# others to follow VideoPose3d, PoseFormer and SwingNet
results
PCE of CBAM-SwingNet in golfdb evaluate
| model | PCE |
|---|---|
| SwingNet | 76.1% (reported in the paper) |
| CBAM-SwingNet | 80.5% (split_4_flip+affine_7700.pth.tar) |
F1-score for start-end detect (1~396 remove defective video)
| Methods | F1-score-aver |
|---|---|
| SwingNet-based | 0.7425 |
| distance_threshold-based | 0.7599 |
| CBAM-SwingNet-based | 0.779 |
Pearson Correlation Coefficient (distance_threshold + dtw based)
| No. | mean | var | pearson | p-value |
|---|---|---|---|---|
| 1~190 | 0.497646 | 0.008536 | -0.21548 | 0.0028 |
| 1~396 | 0.531162 | 0.011472 | -0.15446 | 0.0021 |
For details, see the csv file in the results.
