PersFormer: A New Baseline for 3D Lane Detection
Introduction
PersFormer is a cutting-edge tool designed for the detection of 3D lane lines using a monocular camera. It utilizes a novel Transformer-based feature transformation module, specifically crafted to enhance the detection of lane lines in a three-dimensional space. The core advantage of PersFormer lies in its ability to generate bird's-eye-view (BEV) features, allowing it to focus on pertinent local regions in front-view images with the help of camera parameters.
One of the key differentiators of PersFormer is its unified design, which simultaneously detects 2D and 3D lanes. This concurrent detection process boosts the consistency of features and benefits from the synergy of multi-task learning.
Changelog
Significant updates in the project timeline include the following:
- On November 3, 2022, a bug in the evaluation pipeline was fixed and the optimal version of the PersFormer was made available for the OpenLane dataset.
- On September 27, 2022, evaluation metrics were updated to improve accuracy by considering lane visibility during evaluation.
- On May 9, 2022, a comparison was made with other methods, where PersFormer showed superior results.
Getting Started
Installation
PersFormer requires a machine equipped with at least one GPU. The installation process can be followed through the provided guidelines, ensuring the necessary environment is set up correctly.
Dataset
Data for training and evaluation can be obtained from the OpenLane dataset and the Apollo 3D Lane Synthetic Dataset. Detailed instructions for downloading these datasets are available via the relevant links.
Training and Evaluation
Comprehensive steps to train and evaluate PersFormer are documented, ensuring users can effectively utilize the tool for their specific needs.
Benchmark
PersFormer excels in both 3D and 2D lane detection tasks. Benchmarked against other methods on the OpenLane and ONCE 3DLanes datasets, PersFormer consistently performs better in various challenging environments like curves, extreme weather, and intersections.
3D lane detection performance (measured in F-Score) in OpenLane shows PersFormer achieving superior scores across different categories compared to competitors like GenLaneNet and 3DLaneNet. Furthermore, in 2D lane detection, while using the LaneATT methodology as a baseline, PersFormer outperforms with a noteworthy F-Score. On the ONCE 3DLanes dataset, PersFormer again shows its prowess, leading in precision and recall metrics.
Visualization
Visualization tools are provided for users to observe the results of lane detection using PersFormer. These visualizations give insights into how accurately and efficiently the system detects lane lines in various scenarios and datasets.
Citation
Researchers or developers using PersFormer in their work are encouraged to cite the PersFormer paper for recognition of the efforts and innovation that drove the project.
Acknowledgements
This project acknowledges significant contributions and support from various individuals and organizations, highlighting collaborative efforts with SenseTime Research and constructive feedback from academic and research professionals.
License
The PersFormer project is available under the Apache License 2.0, promoting open-source collaboration while protecting intellectual property rights.