Lane Detection: An Instance Segmentation Approach

Introduction

This is a PyTorch implementation and variation of the paper "Towards End-to-End Lane Detection: An Instance Segmentation Approach".

Basic Results

Results on TuSimple Benchmark test set

Architecture	Accuracy	FP	FN	FPS
FCN-Res18	0.940	0.142	0.085	15.6
FCN-Res34	0.941	0.133	0.083	14.6
ENet	0.937	0.149	0.093	10.8
ICNet	0.935	0.139	0.103	11.1

Note:

• The model was trained using only TuSimple dataset, without data augmentation, the size of input images is 512*288.
• Training and testing were performed on an computer cluster with Intel Xeon CPUs(E5-2690 v4 @ 2.60GHz) and NVIDIA Titan V GPUs.

Quick Start

Preparation

1. Install dependencies:

   pip install -r requirements.txt
   

2. Download TuSimple Benchmark dataset, and unzip the packs. The dataset structure should be as follows:

   tusimple_benchmark
    `-- |-- test_set
        |   |-- clips
        |   `-- ...
        `-- train_set
            |-- clips
            |-- label_data_xxxx.json
            |-- label_data_xxxx.json
            |-- label_data_xxxx.json
            `-- ...
   

3. Download checkpoint pth files from our LaneNet model zoo.

   • OneDrive
   • 百度网盘 (提取码: u6wa)

Training and Testing

Testing on TuSimple Benchmark

python test_lanenet-tusimple_benchmark.py \
        --data_dir /path/to/test_set \
        --arch <MODEL> \
        --ckpt_path /path/to/checkpoint/file

• Setup testing set path using --data_dir /path/to/test_set
• Select network architecture by --arch <MODEL>, options include fcn, enet, icnet, and fcn is the default option.
• Add --dual_decoder to use seperate decoders for the binary segmentation branch and embedding branch. By default, these two branches shares a decoder.
• Checkpoint file should be a *.pth file.
• Add --show to display output images while testing. In each iteration, after show images ,the program pauses until a key is pressed.
• Add --save_img to save images into ./output/ while testing.
• Add --ipm to conduct Inverse Projective Mapping(IPM) before fitting lane curves.
• By passing --tag <string>, one can record experimental settings notices in a string, which will be included in the name of output directories and log files.

Training on TuSimple Benchmark

python train_lanenet.py \
        --data_dir /path/to/train_set \
        --arch <MODEL> \
        --ckpt_path /path/to/checkpoint/file

• Setup training set path using --data_dir /path/to/train_set, both training and validation data are loaded from this directory.
• Select network architecture by --arch <MODEL>, options include fcn, enet, icnet, and fcn is the default option.
• Add --dual_decoder to use seperate decoders for the binary segmentation branch and embedding branch. By default, these two branches shares a decoder.
• Checkpoint file should be a *.pth file.
• By passing --tag <string>, one can record experimental settings notices in a string, which will be included in the name of output directories and log files.

Dataset division and analysis

Architecture

Tensorboard Summary Details

TODO

• integrate TuSimple Bencnmark eval script with the test script
• Discuss about IPM, dataset division.