Hand Gesture Recognition using Small HGR1 Dataset

Motivation

Last year, I visited a charity school in Vietnam for children who are deaf and have disabilities. I wanted to find a way to connect with them, and I thought that sign language could be a solution. While there are many sign language recognition models out there, I decided to create my own. Despite not having access to a powerful PC cluster or a large dataset, I aimed to improve accuracy using a small dataset to see how effective it can be.

This Project

In this project, I use my experience in Computer Vision and Data Science to push the limits of a small dataset and explore how much accuracy can be achieved. I am using ResNet50 for its excellent performance and transfer learning to improve the model's effectiveness. The results are noticeable when comparing the 99% accurate ResNet50 model to the 96% fine-tuned model. The Resnet50 seems to be too generalized and inefficient. However after transfer learning and a little bit of fine-tuning the model works much better!

Conclusion

This project was a great exercise in working with a small dataset. Different approaches resulted in varying levels of accuracy, and the project can be further modified and improved.

Built With

• keras
• python