Light Conditions Style Transfer

Paper

Lane Detection in Low-light Conditions Using an Efficient Data Enhancement : Light Conditions Style Transfer

Submitted to IV 2020

The main framework is as follows:

Empirically, lane detection model trained using our method demonstrated adaptability in low-light conditions and robustness in complex scenarios. (It can achieve 73.9 F1-measure in CULane testing set)

Datasets

CULane

The whole dataset is available at CULane.

CULane
├── driver_23_30frame       # training&validation
├── driver_161_90frame      # training&validation
├── driver_182_30frame      # training&validation
├── driver_193_90frame      # testing
├── driver_100_30frame      # testing
├── driver_37_30frame       # testing
├── laneseg_label_w16       # labels
└── list                    # list

Generated Images

The images in low-light conditions are generated by the proposed Better-CycleGAN. We will upload our generated images in low-light conditions soon.

Source Code

Better-CycleGAN

We will open the source code for Better-CycleGAN soon.

ERFNet

The source code used for the lane detction is made publicly available by HOU Yuenan.

Requirements

PyTorch 0.3.0.
Matlab (for tools/prob2lines), version R2017a or later.
Opencv (for tools/lane_evaluation), version 2.4.8 or later.

Test

The trained model used in this paper is available in ./trained.

Run test script

sh ./test_erfnet.sh

Get lines from probability maps

cd tools/prob2lines
matlab -nodisplay -r "main;exit"

Please check the file path in Matlab code before.

Evaluation

cd /tools/lane_evaluation
make
sh Run.sh # run.sh

Run.sh evaluate each scenario separately while run.sh evaluate the whole. The evaluation results are saved in /tools/lane_evaluation/output

Performance

Light Conditions Stlye Transfer

Some examples of real images in normal light conditions and their corresponding translations images in low-light conditions.

Lane Detetcion

Performance ( (F₁-measure) ) of different methods on CULane testing set. For crossroad, only FP is shown.

Category	ERFNet	CycleGAN+ERFNet	Better-CycleGAN + ERFNet(ours)	SCNN	ENet-SAD	ResNet-101-SAD
Normal	91.5	91.7	91.8	90.6	90.1	90.7
Crowded	71.6	71.5	71.8	69.7	68.8	70.0
Night	67.1	68.9	69.4	66.1	66.0	66.3
No Line	45.1	45.2	46.1	43.4	41.6	43.5
Shadow	71.3	73.1	76.2	66.9	65.9	67.0
Arrow	87.2	87.2	87.8	66.9	65.9	67.0
Dazzle Light	66.0	67.5	66.4	58.5	60.2	59.9
Curve	71.6	69.0	72.4	64.4	65.7	65.7
Crossroad	2199	2402	2346	1990	1998	2052
Total	73.1	73.6	73.9	71.6	70.8	71.8

The probability maps output by the three methods above are shown as following

To do

Upload the generated images
Open the source code for Better-CycleGAN

Acknowledgement

This project refers to the following projects:

wzh7599 / light-condition-style-transfer Goto Github PK

light-condition-style-transfer's Introduction