Hand-Gesture-Recognition

Final Project for CAP 5415: Computer vision

We have used the Google’s high-fidelity hand and finger tracking API, MediaPipe Hands, as feature extractor of hand skeletal features. Images used in this project is collected from a recently published dataset for hand gesture detection, HAGRID.

We have trained multiple Machine Learning models on the extracted features to recognize hand gestures.

Trained Models:

DNN

3 fully connected layers consisting of 256, 128, and 128 neurons followed by ReLU activation and dam optimizer. Learning rate of 0.0005 for optimizing the model over categorical cross-entropy loss.

CNN

2 convolutional layers followed by ReLU activations and 50% dropout. The output is then fed to 3 fully connected layers consisting of 128, 64 and 6 neurons, where the first two FC layers are followed by ReLU activation and the last one by softmax activation.

Downloads

Subsample

Subsample has 100 items per gesture.

Subsample	Archives	Size
images	subsample	2.5 GB
annotations	ann_subsample	1.2 MB

Subsample Processed

Hand landmarks - gesture labeled

Subsample Processed Data

Dataset

Gesture	Size	Gesture	Size
call	39.1 GB	peace	38.6 GB
dislike	38.7 GB	like	38.3 GB
fist	38.0 GB	rock	38.9 GB

Processed

Hand landmarks - gesture labeled

Processed Data

Report can be found here: report.docx