In this project, you'll label the pixels of a road in images using a Fully Convolutional Network (FCN).
Because the implementation in main.py
has passed all the tests so it has met the criteria.
-
The function
load_vgg
is implemented correctly. -
-
The function
layers
is implemented correctly. -
-
The function
optimize
is implemented correctly. -
-
The function
train_nn
is implemented correctly. -
Yes, it does. you can see the loss (%) over time in the following figure:
Loss (%) Over Time |
---|
In addition, during the training you can see the progress over time in terminal. I put the log info in log.txt file.
I chose the following hyper parameters for the network and training:
image_shape = (160, 576)
num_classes = 2
epochs = 20
batch_size = 8
keep_prob_value = 0.75
learning_rate_value = 1e-4
Yes, it does. You can find the labeled images in ./runs/1541284958.1358018. Also I put a fast preview of them below:
Fast Preview of Labeled Images |
---|
The network architecture for the network is as follows:
[vgg3] [vgg4] [vgg7]
| | |
| | \--->[conv2d:1x1]---\
| | |
| [scale] [x] V
| 0.01 | [conv2d transpose] #1
| | |
| | V
| \--->[conv2d:1x1]------------>[+] [add] #1
[x] [scale] |
| 0.0001 V
| [conv2d transpose] #2
| |
| V
\-------------->[conv2d:1x1]------------>[+] [add] #2
|
V
[conv2d transpose] (output layer)
I used the Kitti Road dataset from here. Also I generated new jittered date and add them to the data set to have a better training data set.
To generate the jittered images I used OpenCV to scale up, rotate CW/CCW, and flip them (From line 178 to line 200). Also, I adjust the gamma of images (line 203).
It all helps to bring more variety to data set of images and improve road recognition at the end.
Note
I used Udacity's classroom workspace. So, I set
data_dir
to'/data'
(line 250). If you want to run it somewhere else make sure change it todata_dir = './data'
.
main.py
will check to make sure you are using GPU - if you don't have a GPU on your system, you can use AWS or another cloud computing platform.
Make sure you have the following is installed:
Download the Kitti Road dataset from here. Extract the dataset in the data
folder. This will create the folder data_road
with all the training a test images.
Implement the code in the main.py
module indicated by the "TODO" comments.
The comments indicated with "OPTIONAL" tag are not required to complete.
Run the following command to run the project:
python main.py
Note If running this in Jupyter Notebook system messages, such as those regarding test status, may appear in the terminal rather than the notebook.
- The link for the frozen
VGG16
model is hardcoded intohelper.py
. The model can be found here. - The model is not vanilla
VGG16
, but a fully convolutional version, which already contains the 1x1 convolutions to replace the fully connected layers. Please see this post for more information. A summary of additional points, follow. - The original FCN-8s was trained in stages. The authors later uploaded a version that was trained all at once to their GitHub repo. The version in the GitHub repo has one important difference: The outputs of pooling layers 3 and 4 are scaled before they are fed into the 1x1 convolutions. As a result, some students have found that the model learns much better with the scaling layers included. The model may not converge substantially faster, but may reach a higher IoU and accuracy.
- When adding l2-regularization, setting a regularizer in the arguments of the
tf.layers
is not enough. Regularization loss terms must be manually added to your loss function. otherwise regularization is not implemented.