SMERF: Augmenting Lane Perception and Topology Understanding with Standard Definition Navigation Maps

This is the official code release for

Augmenting Lane Perception and Topology Understanding with Standard Definition Navigation Maps.

by Katie Z Luo, Xinshuo Weng, Yan Wang, Shuang Wu, Jie Li, Kilian Q. Weinberger, Yue Wang, and Marco Pavone

Paper | Project Page

Abstract

Autonomous driving has traditionally relied heavily on costly and labor-intensive High Definition (HD) maps, hindering scalability. In contrast, Standard Definition (SD) maps are more affordable and have worldwide coverage, offering a scalable alternative. In this work, we systematically explore the effect of SD maps for real-time lane-topology understanding. We propose a novel framework to integrate SD maps into online map prediction and propose a Transformer-based encoder, SD Map Encoder Representations from transFormers, to leverage priors in SD maps for the lane-topology prediction task. This enhancement consistently and significantly boosts (by up to 60%) lane detection and topology prediction on current state-of-the-art online map prediction methods without bells and whistles and can be immediately incorporated into any Transformer-based lane-topology method.

Table of Contents

• Main Results
• Installation
• Dataset Setup
• Train and Evaluate
• Citation

Main Results

Results on OpenLane-V2 subset-A val

We provide results on Openlane-V2 subset-A val set.

	DETl	TOPll	DETt	TOPlt	OLS	Model	Config
Baseline	17.0	2.3	48.5	16.2	30.2	ckpt	cfg
Baseline + SMERF	26.8	3.9	48.9	19.2	34.8	ckpt	cfg
Toponet	28.2	4.1	44.5	20.6	34.5	ckpt	cfg
Toponet + SMERF	33.4	7.5	48.6	23.4	39.4	ckpt	cfg

Installation

Prerequisites

• Linux
• Python 3.8.15
• NVIDIA GPU + CUDA 11.1
• PyTorch 1.9.1

Environment Setup

We recommend using conda to run the code. Alternatively, we provide a Dockerfile for ease of installation.

conda create -n smerf python=3.8 -y
conda activate smerf

pip install torch==1.9.1+cu111 torchvision==0.10.1+cu111 -f https://download.pytorch.org/whl/torch_stable.html

Install mm-series packages.

pip install mmcv-full==1.5.2 -f https://download.openmmlab.com/mmcv/dist/cu111/torch1.9.0/index.html
pip install mmdet==2.26.0
pip install mmsegmentation==0.29.1

Install other required packages.

cd SMERF
pip install -r requirements.txt

This code base uses mmdetection3D, and all relevant code for the project is in the mmdetection3d/projects directory. Set up mmdetection3D, corresponding to mmdet3d==1.0.0rc6:

cd SMERF/mmdetection3d
pip install -v -e .

Install dependencies for OpenStreetMap:

pip install osmnx==1.5.1
pip install av2==0.2.1
pip install geocube==0.3.3
pip install sparse==0.14.0

If errors occur, manually upgrade the following packages:

pip install --upgrade jupyter_core jupyter_client
pip install --upgrade "protobuf<=3.20.1"

Dataset Setup

Follow the OpenLane-V2 repo to download the data. Preprocess the data for training and evaluation:

cd data
python OpenLane-V2/preprocess.py 

After setup, the hierarchy of folder data/OpenLane-V2 should look as follows:

data/OpenLane-V2
├── train
|   └── ...
├── val
|   └── ...
├── test
|   └── ...
├── data_dict_subset_A_train.pkl
├── data_dict_subset_A_val.pkl
├── data_dict_subset_B_train.pkl
├── data_dict_subset_B_val.pkl
├── ...

SD Map Processing

This work uses Standard Definition (SD) maps; our SD maps are pulled from OpenStreetMap. Download and process the SD maps relevant to this dataset:

cd openlanev2
python sd_maps/load_sdmap_graph.py --collection data_dict_subset_A_[train/val] --city_names [train/val]_city

A script for parallelizing the process is provided at openlanev2/sd_maps/load_sdmap.sh.

Train and Evaluate

Train

This work reported numbers for models trained with 8 GPUs. If a different number of GPUs is utilized, you can enhance performance by configuring the --autoscale-lr option.

cd SMERF/mmdetection3d
bash tools/dist_train.sh [config] 8 [--autoscale-lr]

The training logs will be saved to work_dirs/[config].

For example, to train the OpenLane-V2 baseline model with SMERF map embeddings on 8 GPUs, run:

bash tools/dist_train.sh projects/openlanev2/configs/baseline_large_ptsrep_smerf.py 8

Edit: It has been brought to my attention that the baseline model might take too much memory for most machines. One thing I would suggest is to run the Toponet baseline. Their model is more lightweight and (as a plus) yields better results. Run distributed training with:

bash tools/dist_train.sh projects/toponet_openlanev2/configs/toponet_smerf.py 8

Evaluation and Visualization

You can set --eval-options visualization=True to visualize the results.

cd SMERF/mmdetection3d
bash tools/dist_test.sh [config] [checkpoint_path] 8 --eval * [--eval-options visualization=True visualization_num=200]

Training and evaluating on geographically disjoint set

To reproduce the split from the geographically disjoint OpenLane-V2 split, run the preprocessing code:

cd data
python OpenLane-V2/process_disjoint_split.py

The model configs and checkpoints are provided for training and evaluation:

	OLS	Model	Config
Baseline	16.9	ckpt	cfg
Baseline + SMERF	22.1	ckpt	cfg
Toponet	21.7	ckpt	cfg
Toponet + SMERF	23.4	ckpt	cfg

Citation

If this work is helpful for your research, please consider citing us!

@article{luo2023augmenting,
  title={Augmenting Lane Perception and Topology Understanding with Standard Definition Navigation Maps},
  author={Luo, Katie Z and Weng, Xinshuo and Wang, Yan and Wu, Shuang and Li, Jie and Weinberger, Kilian Q and Wang, Yue and Pavone, Marco},
  journal={arXiv preprint arXiv:2311.04079},
  year={2023}
}