SceneNet | Devpost

Inspiration

Empowering those in need with technology, along with examples from our lives.

What it does

Camera on glasses used to recognize scenes using neural network to help aide those who are visually impaired in understanding their surroundings.

How we built it

We used PyTorch to train our neural network using an MIT database of over 2.5 million images to recognize scenes. We then offloaded the processing to a Jetson TX2 for portability.

Challenges we ran into

Dependencies. Mainly training the network to run on an ARM processor instead of x86 architecture.

Accomplishments that we're proud of

Getting the network to classify images.

What we learned

ARM processor architecture is way different than x86.

What's next for SceneNet

Add localization and be able to extended the use cases to classify more dynamic scenes and objects.

Built With

cuda
cufnn
jetson
python
pytorch
ubuntu

Submitted to

HackPrinceton Spring 2018

Created by

Helped with neural network design and accelerating inference with tensorRT and cuda. This was my first time deploying a neural network for a practical use case and I learned a ton. Especially about different ways to improve inference and overall accuracy. I also modeled the design for the Jetson belt clip holder in autoCAD.

Khalid Khalil
Helped with bridge between hardware and software, especially with the implementation of the TTS. Also with the deployment of the TX2 from x86 to ARM.

Muhammad Khalil
ayman-saleh Saleh

Updates

ayman-saleh Saleh started this project — Apr 01, 2018 08:42 AM EDT

Leave feedback in the comments!

Log in or sign up for Devpost to join the conversation.