• ⚡⚡ Super lightweight: The number of model parameters is 267,867 (< LSTR's 765,787).
• ⚡⚡ Super low complexity: The number of MACs (1 MAC = 2 FLOP) is only 486.617M (< LSTR's 574.280M).
• 😎 Learning structures with large curvatures without expensive, dense, and precise human annotations (predicting curved lanes better than LSTR).

The LLAMAS dataset paper defined and required lane categories when evaluating algorithms. l0 means the closest left lane to the host-vehicle, r0 means the closest right lane (likewise for l1 and r1).

In these demo videos, no tracking module is used.

The LI is trained without the proposed curvature-aware learning.

The pretrained models are stored in LICAZoos/

• Linux ubuntu 16.04
• GeForce RTX 3090
• Python 3.8.5
• CUDA 11.1

Create virtualenv environment

python3 -m venv lica

Activate it

source lica/bin/activate

Then install dependencies

pip install torch==1.8.1+cu111 torchvision==0.9.1+cu111 torchaudio==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
pip install -r requirements.txt

Download and extract TuSimple, and LLAMAS.

We expect the directory structure to be the following:

lica/
LICA/
    LICA/
    LICAZoos/
TuSimple/
    LaneDetection/
        clips/
        label_data_0313.json
        label_data_0531.json
        label_data_0601.json
        test_label.json
LLAMAS/
    color_images/
    labels/

python test.py LICA_TR2_TUSIMPLE --testiter 500000

python test.py LICA_TR6_TUSIMPLE --testiter 500000

python test.py LICA_TR2_LLAMAS --testiter 500000 --split validation

Corresponding codes will be released after acceptance.

LSTR