gongyan1 / TCLaneNet

🎉 [TIV' 2024]TCLaneNet: Task-conditioned Lane Detection Network Driven by Vibration Information

TCLaneNet Pytorch

Pytorch implementation of "TCLaneNet: Task-conditioned Lane Detection Network Driven by Vibration Information"

Changelog

[2022-2-13] Release the initial code for TCLaneNet.

一、Demo

1. Video

Demo trained with VBLane Dataset.

2. Comparison

The visualization results compared with other methods, where red and green represent false positive and false negative respectively.

二、Train and test

1. Requirements

• pytorch
• opencv
• numpy

2. Datasets

• VBLane Dataset

  VBLane dataset path (click to expand)

  VBLane_path
  ├─ no
    ├─ ano_heng_1-2021-04-12-18-45-23
    ├─ ano_heng_2-2021-04-12-18-47-00
    ├─ ano_heng_3-2021-04-12-18-48-28
    ├─ ano_road_1-2021-04-12-18-37-12
    ├─ ano_road_2-2021-04-12-18-39-02
    ├─ ano_road_3-2021-04-12-18-40-09
  ├─ yes
    ├─ new2_road_again_2-2021-04-12-17-52-27
    ├─ new2_road_again_3-2021-04-12-17-53-30
    ├─ new2_road_again_4-2021-04-12-17-54-56
    ├─ new2_road_again_5-2021-04-12-17-56-39
    ├─ new2_road_again_6-2021-04-12-17-58-01
  └─ list
    ├─ train.txt
    ├─ test.txt
  

You need to change the correct dataset path in ./config.py

Dataset_Path = dict(
    CULane = "/workspace/CULANE_DATASET",
    Tusimple = "/workspace/TUSIMPLE_DATASET",
    bdd100k = "/workspace/BDD100K_DATASET",
    mydata = "/home/neu-wang/gongyan/big_data/vibration"
)

3. Training

First, change some hyperparameters in ./experiments/*/cfg.json

{
  "model": "TCLaneNet", 
  # Optional: "SCNN", "LaneNet", "Enet", "Enet-SAD", "Resa", or "TCLaneNet"

  "dataset": {
    "dataset_name": "mydata", 
    "batch_size": 12,
    "resize_shape": [800, 288]       
    #[800, 288] with CULane and VBLane, [640, 368] with Tusimple, and [640, 360] with BDD100K
  },
  ...
}

And then, start training with train.py.

python train.py --exp_dir ./experiments/TCLaneNet

If the training is interrupted, the following command is used to resume.

python train.py --exp_dir ./experiments/lanenet --resume

4. Testing

First, change some hyperparameters in ./experiments/*/cfg.json, e.g., model_path, test.

"evaluate": {
    "test": "yes",
    "model_path": "./experiments/TCLaneNet/TCLaneNet_iou_best.pth"
  }

Then, start testing with test.py.

python test.py --exp_dir ./experiments/TCLaneNet

What's more, if the test is yes, you can python train.py --exp_dir ./experiments/TCLaneNet.

三、Performance

• VBLane dataset

Method	Rec	Pre	F1	ACC	IoU
LaneNet	97.67	95.28	96.46	99.41	93.18
SCNN	93.34	97.32	95.28	99.21	91.02
RESA	96.81	95.56	96.17	99.38	92.65
ENet	92.53	97.85	95.11	99.18	90.69
ENet-SAD	94.62	96.65	95.62	99.27	91.62
ERFNet	97.64	92.86	95.25	99.20	91.21
ERFNet-HVESA	96.65	94.78	95.64	99.31	91.82
TCLaneNet (ours)	98.53	95.64	97.06	99.51	94.29

Compared with other methods, our method has better performance in F1, IoU and ACC.

Acknowledgement

This repo is built based on ENet-SAD, ENet-SAD-Pytorch, PyTorch-ENet, and SCNN_Pytorch.