lyk125 / udacity-driverless-car-nd-p2 Goto Github PK

# udacity-driverless-car-nd-p2

This is my work for the Self-Driving Car Engineer Nanodegree ND-013 course project, "2.Traffic Sign Classifier" . The project problem can be found here: https://github.com/udacity/CarND-Traffic-Sign-Classifier-Project

Please see the "basic" folder.

Let us look at the data (the second column is the mean image of each class):

Here is the basic solution using LeNet and Vgg.

Please see the "advance" folder.

There is some write up at my blog:

https://medium.com/@hengcherkeng/updated-my-99-40-solution-to-udacity-nanodegree-project-p2-traffic-sign-classification

https://medium.com/@hengcherkeng/updated-my-99-68-solution-to-udacity-nanodegree-project-p2-traffic-sign-classification

I use modified densenet[1] and obtained 99.68% on the test set. The network complexity is about 27.0 million MAC (multiply–accumulate operation counts).

Here is my network structure. Each "Dense block" consists of concatenation of convolutions (in conv-bn-relu). Note that unlike the paper, dropout is not applied in the block. Instead, I use droupout after the block.

Here is the MAC computation

Finally, the loss curves are shown below.

[1] "Densely Connected Convolutional Networks" - Gao Huang, Zhuang Liu, Kilian Q. Weinberger, Laurens van der Maaten, Arxiv 2016