This is a Pi-based robot to implement visual recognition(by Inception V3). The TensorFlow-Powered vision can recognize many objects such as people, car, bus, fruits, and so on.
-
Hardware: Raspberry-Pi2, Sony PS3 Eye Camera
(Available to use Logitech C270 USB camera with Raspberry Pi)
-
Software: TensorFlow(v1.0.1), Jupyter-Notebook
I was so curious about excellence of the image recognition with TensorFlow on Raspberry Pi. Also, the Jupyter notebook is very convenient to instantly code as a quick prototype. So, in terms of error rate of the image classification, Inception V3(3.46%) is more excellent than human(5.1%) whereas raspberry pi's processing speed is very slow compare to my laptop.
(Chart: Jeffrey Dean's Keynote @Google Brain).
- Schematic diagram of Inception-v3
-
TensorFlow (V1.0.1): How To Install TensorFlow on Raspberry Pi
-
Jupyter-Notebook: How To Install Jupyter-Notebook on Raspberry Pi
- You should install both TensorFlow(v1.0.1) and Jupyter-Notebook on your Raspberry Pi.
- First, clone the TensorFlow-Powered_Robot_Vision git repository on your Raspberry Pi. This can be accomplished by:
cd /home/pi/Documents
git clone https://github.com/leehaesung/TensorFlow-Powered_Robot_Vision.git
Next, cd into the newly created directory:
cd TensorFlow-Powered_Robot_Vision
Let's start Jupyter-notebook on cammand window.
jupyter-notebook
When you implement the jupyter-notebook, the trained data(inception_v3.ckpt) will automatically download there.(/pi/home/Documents/datasets/inception)
- TensorFlow-Powered_Robot_Vision.ipynb
- imagenet_class_names.txt
- inception_v3.ckpt (When implementing Jupyter notebook)
- Wow! This result is really awessome!!
- Rethinking the Inception Architecture for Computer Vision (Paper)
- Image Recognition::Tensorflow.org (Web)
- Hands-On Machine Learning with Scikit-Learn and TensorFlow (Book)
- Train your own image classifier with Inception in TensorFlow (Blog)
- "Large-Scale Deep Learning for Building Intelligent Computer Systems," a Keynote Presentation from Google