README.md

On-Device-Deep Learning

Files used:

Initial Training

• trainer.py was used to train our model initially
• inference.py was used to test the model

Revised Training

• test_gesture.py is used to gather landmark coordinates for sequences of images using Mediapipe Hand Landmarker
  • see https://developers.google.com/mediapipe/solutions/vision/hand_landmarker for more details
• prepare_data.py is used to convert the jester dataset image sequences into landmarks returned in test_gesture.py
• model.py is used to train a LSTM model on the prepared data
  • model is saved to a checkpoint every 5 epochs
  • model history is saved to trainHistoryDict
• Evaluate_models.ipynb is used to load model history and visualize/benchmark the performance
  • bella_model.h5 is saved after the training is complete
  • pruned_and_quantized.tflite is then created with a signature specified to enable use in the tflite runtime

On-Device

• test_tflite.py is used to test the accuracy of the tensorflow lite model on Jester test data
• single_gesture_classifier.py is used to collect a 3 single second gesture and then perform classification
• real_time_classifier.py is used to perform continuous gesture classification in real-time

Performance(Inference.py) after initial training using trainer.py

Demo Videos

Performing asynchronous gesture classification, using the file 'Test_tflite.py':

swipe_down.mp4

Performing real-time gesture classification, using the file 'Test_gesture.py':

PXL_20231215_005556192.3.mp4

IMG_2368.online-video-cutter.com.1.mov

Authorship

Shivam Sharma: trainer.py, inference.py

Isabella Feeney: test_gesture.py, prepare_data.py, model.py, Evaluate_models.ipynb, test_tflite.py, single_gesture_classifier.py, real_time_classifier.py