Sign Language Digits Recognition

This project was done as the final project for the Fundamentals Of Data Science And Lab class of Sapienza's Data Science's Master Degree.

Team

• Valentino Sacco: S4b3
• Arturo Ghinassi: ghinassi1863151
• Camilla Savarese: Camillasavarese
• Giorgia Fontana: GiorgiaFontana
• Luca Romani: LucaRomani98

Goal

For this project we tried to understand how to best perform recognition on the Sign Language Digits Dataset*, starting from a Naive Bayes Classifier approach and then seeing how to enhance its performances.

Report

An more in depth Report of our work process, observations and results can be found here

Techniques (main.ipynb)

The sequent approaches were tried in combination with dataset transformation techniques to find the best solution

Naive Bayes

• Multinomial Approach

  • Preprocessing of the dataset into black and white pixels as features
    
    

  • Pillow Edge Recognition
    
    

• Gaussian Approach

  • Standard Dataset

  • Principal Component Analysis for noise reduction

Chamfer Distance

to implement the chamfer distance we took this implementation as reference
We computed chamfer distance after applying above edge recognition on the dataset. This is worth mentioning but the results have not lived to our expectations.

Support Vector Machine

• Standard Dataset

• Principal Component Analysis for noise reduction

• Gaussian Smoothing for noise reduction

  

Repository content

• main.ipynb: is the main working area, containing all the

• utils.py: module containing data loading and preprocessing functions

• naive_bayes_custom.py: module containg a custom implementation of the Naive Bayes Classifier used for error understanding

• X.npy: dataset images in the form of a numpy matrix

• Y.npy: numpy vector containing true labels of the images

*"Mavi, A., (2020), “A New Dataset and Proposed Convolutional Neural Network Architecture for Classification of American Sign Language Digits”, arXiv:2011.08927 [cs.CV]"*