This classifier has an architecture inspired by LeNet.
| Layer | Description |
|---|---|
| Input | 32x32x1 grayscale image |
| Convolution 5x5 | 1x1 stride, valid padding, outputs 28x28x6 |
| RELU | |
| Max pooling | 2x2 stride, outputs 14x14x6 |
| Convolution 5x5 | 1x1 stride, valid padding, outputs 10x10x16 |
| RELU | |
| Max pooling | 2x2 stride, outputs 5x5x16 |
| Fully connected | outputs 120 |
| RELU | |
| Dropout | 50% keep probability |
| Fully connected | outputs 84 |
| RELU | |
| Fully connected | outputs 43 |
My final model results were:
- training set accuracy of 0.997
- validation set accuracy of 0.973
- test set accuracy of 0.955
I have run the classifier on these images found online.
| image 0 (Yield) | image 1 (Turn right ahead) | image 2 (Priority road) | image 3 (Speed limit 30km/h) | image 4 (Speed limit 120km/h) | image 5 (No passing) |
|---|---|---|---|---|---|
 |
 |
 |
 |
 |
 |
Here are the results of the prediction:
| Image | Prediction |
|---|---|
| Yield | Yield |
| Turn right ahead | Turn right ahead |
| Priority road | Priority road |
| Speed limit 30km/h | Speed limit 30km/h |
| Speed limit 120km/h | Speed limit 120km/h |
| No passing | No passing |
The model was able to correctly guess 6 of the 6 traffic signs, which gives an accuracy of 100%. This compares favorably to the accuracy on the test set of accuracy 95.5%.
These are the top five probabilities of predictions on each emage
image 0
| Probability | Prediction |
|---|---|
| 1.00 (lost precision) | Yield |
| 8.28896e-14 | No vehicles |
| 5.02711e-15 | Priority road |
| 4.1249e-15 | Stop |
| 7.78334e-16 | Keep right |
image 1
| Probability | Prediction |
|---|---|
| 0.999994 | Turn right ahead |
| 5.67283e-06 | Ahead only |
| 6.09594e-10 | No vehicles |
| 4.83902e-10 | Keep left |
| 2.11305e-10 | Roundabout mandatory |
image 2
| Probability | Prediction |
|---|---|
| 1.0 (lost precision) | Priority road |
| 8.71962e-09 | Stop |
| 2.33766e-09 | No vehicles |
| 7.9627e-10 | End of all speed and passing limits |
| 3.52491e-10 | No passing |
image 3
| Probability | Prediction |
|---|---|
| 0.998181 | Speed limit (30km/h) |
| 0.0010598 | Speed limit (80km/h) |
| 0.000508322 | Speed limit (50km/h) |
| 0.000196236 | End of speed limit (80km/h) |
| 3.89811e-05 | Speed limit (20km/h) |
image 4
| Probability | Prediction |
|---|---|
| 0.7852 | Speed limit (120km/h) |
| 0.10838 | Speed limit (70km/h) |
| 0.10636 | Speed limit (20km/h) |
| 4.01502e-05 | Speed limit (80km/h) |
| 1.75141e-05 | Speed limit (100km/h) |
Even though the prediction is correct on image 4, the model considers a little bit of possibility that the sign could be Speed limit (70km/h) or Speed limit (20km/h).
image 5
| Probability | Prediction |
|---|---|
| 0.999999 | No passing |
| 7.57351e-07 | Slippery road |
| 1.13976e-07 | No passing for vehicles over 3.5 metric tons |
| 1.96753e-08 | End of no passing |
| 4.60882e-09 | Dangerous curve to the right |
This program requires:
- Clone the project, which contains the data.
git clone https://github.com/udacity/CarND-Traffic-Sign-Classifier-Project
cd CarND-Traffic-Sign-Classifier-Project
jupyter notebook Traffic_Sign_Classifier.ipynb