LANE DETECTION

This is a light weight Tensorflow implementation of lane detection using fully connected convolutional (FCN) network.

A Demo Picture

This is a forked work from link. The implementation is largely based on the reference code provided by the authors of the paper link.

Hightlighted Work

• Take input images of any size.
• Tailored model for performance: 20 layers lighter than pretrained model. 86% less parameter memory.

Prerequisites

• The code was written and tested with tensorflow0.14 and python2.7.
• Dataset is using KITTI Vision Benchmark Suite. Download the training data in Data_zoo/kitti_road/
• Used pretrained model of IMAGENET VGG 19

Run

To train:

python FCN.py --batch_size=5 --learning_rate=1e-5

To evaluate on a random batch in validation set:

python FCN.py --batch_size=5 --mode=visualize

To infer a image:

python FCN.py --batch_size=5 --mode=infer --data_dir=<image directory or path>

• debug flag can be set during training to add information regarding activations, gradients, variables etc.
• Use IPython notebook can be to view results.

Results

After 1000 batches (~200 epochs), on validation set gets:

ACC (pixcel accuracy): 0.963042, mIOU: 0.797466, time: 5 images / 1.8s

Here's the training curves:

Next Works

• Post processing the predicted images. Consider use CRF.
• KITTI dataset is limited (300 images for train). More data is required to learn rich & generalized features and take care of more scene type. Consider use tusimple or synthia.
• Benchmark on public dataset.

Others

• Implementation Paper: Fully Convolutional Networks for Semantic Segmentation (FCNs).
• Video of the presentaion given by the authors on the paper - link
• The model is originally applied on the Scene Parsing Challenge dataset provided by MIT http://sceneparsing.csail.mit.edu/. Take a look at the code to switch dataset.