Please refer to: Self-Driving Car Engineer Nanodegree - Udacity, Term 1 - Project 2
This project uses a deep convolutional neural network to classify traffic signs. Lenet is used as model. The CNN model takes as input the German Traffic Sign Dataset and understands road signs from images. Find another German Traffic Sign Dataset that can be used.
Goto notebook which shows the code as well full results. The notebook also consist result on on images in folder new_images
To run this project, you need Miniconda installed(Follow instructions to quickly install)
To create an environment for this project use the following command:
conda env create -f environment.yml
After the environment is created, it needs to be activated with the command:
source activate carnd-term1
and open the project's notebook P1.ipynb inside jupyter notebook:
jupyter notebook P1.ipynb
I used numpy and matplotlib libraries to get summary of the data set provided. signs data set:
- The size of training set is 34799
- The size of the validation set is 4410
- The size of test set is 12630
- The shape of a traffic sign image is (32, 32, 3)
- The number of unique classes/labels in the data set is 43
My final model consisted of the following layers:
| Layer | Description |
|---|---|
| Input | 32x32x1 Grayscale image |
| Convolution 5x5 | 1x1 stride, Valid padding, outputs 28x28x6 |
| RELU | Activation function |
| Max pooling | 2x2 stride, outputs 14x14x6 |
| Convolution 3x3 | etc. |
| RELU | Activation function |
| Flatten Input | output 1x400 |
| Fully connected | input 1x400 output 120 |
| RELU | Avtivation function |
| Drop Out | keep_prob=0.8 |
| Fully connected | input 1x120 output 84 |
| Max pooling | 2x2 stride, outputs 5x5x16 |
| RELU | Avtivation function |
| Drop Out | keep_prob=0.6 |
| Softmax | Probabilistic Distribution |
My final model results were:
- training set accuracy of 99.5%
- validation set accuracy of 96%
- test set accuracy of 94.3%
For prediction result analysis on images downloaded from net visit writeup.md.