The website for this project can be found at: https://conductor-gt-f2019.github.io/conductortracking/
For each time signature, we recorded approximately 15 minutes of footage (480p, 30fps) of ourselves performing the corresponding gesture.
We then used OpenPose (https://github.com/CMU-Perceptual-Computing-Lab/openpose) to extract the x and y coordinates of our wrists in each video. The script we used for coordinate extraction can be found in get_pose_data_from_video.py.
Using verify_segment_time.py, we annotated the start of each gesture loop. We then processed the annotated data using accumulator.py to produce accumulator matrices containing a mapping of the gesture's points. The accumulator matrices were saved as images to the data folder for use in the training and validation datasets. 75% of each gesture were used for training, and the remaining 25% were used for validation.
| Gesture | # training | # validation | total |
|---|---|---|---|
| 2/2 | 354 | 119 | 473 |
| 3/4 | 296 | 100 | 396 |
| 4/4 | 250 | 84 | 334 |
Link to data: https://drive.google.com/drive/folders/15A89ZKPJKIkJ8ThkmoGz7r9pBUl_xM1a?usp=sharing
https://pytorch.org/tutorials/beginner/transfer_learning_tutorial.html