Contrastive Learning for Lane Detection via Cross-Similarity

Contrastive Learning for Lane Detection via Cross-Similarity

Overview of CLLD

Contrastive Learning for Lane Detection via cross-similarity (CLLD), is a self-supervised learning method that tackles this challenge by enhancing lane detection models’ resilience to real-world conditions that cause lane low visibility. CLLD is a novel multitask contrastive learning that trains lane detection approaches to detect lane markings even in low visible situations by integrating local feature contrastive learning (CL) with our new proposed operation cross-similarity. To ease of understanding some details are listed in the following:

• CLLD employs similarity learning to improve the performance of deep neural networks in lane detection, particularly in challenging scenarios.
• The approach aims to enhance the knowledge base of neural networks used in lane detection.
• Our experiments were carried out using ImageNet as a pretraining dataset. We employed pioneering lane detection models like RESA, CLRNet, and UNet, to evaluate the impact of our approach on model performances.

CLLD architecture

Get started

1. Clone the repository

   git clone https://github.com/sabadijou/clld_official.git
   

   We call this directory as $RESA_ROOT

2. Create an environment and activate it (We've used conda. but it is optional)

   conda create -n clld python=3.9 -y
   conda activate clld

3. Install dependencies

   # Install pytorch firstly, the cudatoolkit version should be same in your system. (you can also use pip to install pytorch and torchvision)
   conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch
     
   # Install kornia and einops
   pip install kornia
   pip install einops
   
   # Install other dependencies
   pip install -r requirements.txt

How to Run CLLD

We conducted pretraining using the training data from ImageNet. However, you are free to utilize other datasets and configurations as needed. The configuration file for our approach can be found in the configs folder.

Once the dataset and new configurations are in place, you can execute the approach using the following command:

python main.py --dataset_path /Imagenet/train --encoder resnet50 --alpha 1 --batch_size 1024 --world_size 1 --gpus_id 0 1 

The following is a quick guide on arguments:

• dataset_path: Path to training data directory
• encoder: Select an encoder for training. resnet18, resnet34, resnet50, resnet101, resnet152, resnext50_32x4d,resnext101_32x8d, wide_resnet50_2, wide_resnet101_2.
• alpha: Cross similarity window size
• batch_size: Select a batch size that suits the GPU infrastructure you are using.
• world_size: For example, if you are training a model on a single machine with 4 GPUs, the world size is 4. If you have 2 machines, each with 4 GPUs, and you use all of them for training, the world size would be 8.
• gpus_id: Please specify all the GPU IDs that you used for training the approach.

How to publish weights

Upon completing the training phase, you can execute the command below to prepare the trained weights for use as prior knowledge in the backbone of a lane detection model.

python main.py --checkpoint path/to/checkpoint --encoder resnet50 

Our experiments

We specifically chose to evaluate CLLD with U-Net because it is a common encoder-decoder architecture used in various methods that approach lane detection as a segmentation-based problem. In addition, we tested our method using RESA, which is currently the state-of-the-art semantic segmentation lane detection method that is not based on the UNet architecture.This independent validation is necessary to ensure the accuracy of our results. Lastly, we evaluated CLLD using CLRNet, a leading anchor-based lane detection method.

Visualized results

Performance of UNet on CuLane and TuSimple with different contrastive learnings.

Method	# Epoch	Precision (CuLane)	Recall (CuLane)	F1-measure (CuLane)	Accuracy (TuSimple)
PixPro	100	73.68	67.15	70.27	95.92
VICRegL	300	67.75	63.43	65.54	93.58
DenseCL	200	63.8	58.4	60.98	96.13
MoCo-V2	200	63.08	57.74	60.29	96.04
CLLD (α=1)	100	71.98	69.2	70.56	95.9
CLLD (α=2)	100	70.69	69.36	70.02	95.98
CLLD (α=3)	100	71.31	69.59	70.43	96.17

Performance of RESA on CuLane and TuSimple with different contrastive learnings.

Method	# Epoch	Precision (CuLane)	Recall (CuLane)	F1-measure (CuLane)	Accuracy (TuSimple)
PixPro	100	77.41	73.69	75.51	96.6
VICRegL	300	76.27	69.58	72.77	96.18
DenseCL	200	77.67	73.51	75.53	96.28
MoCo-V2	200	78.12	73.36	75.66	96.56
CLLD (α=1)	100	79.01	72.99	75.88	96.74
CLLD (α=2)	100	78	73.45	75.66	96.78
CLLD (α=3)	100	78.34	74.29	76.26	96.81

Performance of CLRNet on CLRNet and TuSimple with different contrastive learnings.

Method	# Epoch	Precision (CuLane)	Recall (CuLane)	F1-measure (CuLane)	Accuracy (TuSimple)
PixPro	100	89.19	70.39	78.67	93.88
VICRegL	300	87.72	71.15	78.72	89.01
DenseCL	200	88.07	69.67	77.8	85.15
MoCo-V2	200	88.91	71.02	78.96	93.87
CLLD (α=1)	100	88.72	71.33	79.09	90.68
CLLD (α=2)	100	87.95	71.44	78.84	93.48
CLLD (α=3)	100	88.59	71.73	79.27	94.25

Acknowledgement

• RESA
• CLRNet
• UNet