Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
End-to-end Lane Detection for Self-Driving Cars (ICCV 2019 Workshop)
| Date | Stars |
|---|---|
| 2026-07-30 | 664 |
| 2026-09-03 | 664 |
| 2026-09-08 | 663 |
| 2026-09-09 | 662 |
| 2026-09-10 | 663 |
| 2026-10-02 | 664 |
| 2026-10-06 | 664 |
| 2026-10-08 | 663 |
| 2026-10-12 | 663 |
| 2026-10-14 | 664 |
| 2026-10-25 | 664 |
| 2026-10-26 | 665 |
| 2026-11-04 | 665 |
Today
— stars today
This week
— stars this week
This month
+1 stars this month
Momentum
0.0
growth rate 0.00%/day
# End-to-end Lane Detection
This repo contains the implementation of our paper [End-to-end Lane Detection through Differentiable Least-Squares Fitting](https://arxiv.org/abs/1902.00293) by Wouter Van Gansbeke, Bert De Brabandere, Davy Neven, Marc Proesmans and Luc Van Gool.
If you find this interesting or relevant for your work, consider citing:
```
@article{wvangansbeke_2019,
title={End-to-end Lane Detection through Differentiable Least-Squares Fitting},
author={Van Gansbeke, Wouter and De Brabandere, Bert and Neven, Davy and Proesmans, Marc and Van Gool, Luc},
journal={arXiv preprint arXiv:1902.00293},
year={2019}
}
```
## Update
I added a new directory __Backprojection_loss__ which is very similar to the other one. However, now the loss is a regression towards the coordinates in the original perspective instead of a regression in the birds eye view perspective towards the lane-line coefficients. We are primarily interested in the accuracy in this perspecitve after all. It also contains multi-lane detection experiments on the complete dataset of TuSimple (3626 images).
## License
This software is released under a creative commons license which allows for personal and research use only. For a commercial license please contact the authors. You can view a license summary [here](http://creativecommons.org/licenses/by-nc/4.0/).
## Setup
This repository compares two methods to achieve higher accuracy for lane detection applications. The former is the conventional segmentation approach and the latter will tackle this problem in an end-to-end manner. The segmentation approach depends on the cross-entropy loss in order to learn the road markings by attention. However this approach is not necessarily the most accurate. Since the final line coordinates are desired, a complete end-to-end method should achieve better results.
This end-to-end architecture consist of two parts. The first part is an off the shelf network to predict weight maps. These weights are applied to a mesh grid which are used for the last step. The last step can be considered as the final layer of the network which solves a weighted system of equations to calculate the final curve parameters of the road markings. The amount of weight maps depends on the amount of lane lines the network ought to detect. This means that a direct regression can be performed to the desired curve coordinates. After all, backpropagation through the whole architecture is now possible.
Finally we show results for egolane detection. The implementation proofs that this direct optimization is indeed possible and can achieve better results than the conventional segmentation approach. Moreover, this module can be applied to a broad range of tasks since the first stage of the architecture can be chosen freely. The extention to other domains such as object detection is considered as future work which also shows the effectiveness of this architecture.
## Requirements
I just updated the code to the most recent version of Pytorch (=pytorch 1.1) with python 3.7.
The other required packages are: opencv, scikit-learn, torchvision, numpy, matplotlib, json/ujson and pillow.
## Dataset
For the egolane detection experiment, I used a subset from the [TuSimple](http://benchmark.tusimple.ai/#/) dataset. You can download it [here](https://drive.google.com/drive/folders/1UECiIOGjIua9ORIDfcZft8XGTQ-iTzuD?usp=sharing). You can find the images as well as the ground truth annotations in this folder to save you some work. Automatically 20% of the data will be used for validation. TuSimple provides the user 3 json files with ground truth coordinates of the lane lines. I appended the three files and used the index in this json file to name a specific image. For example: file "10.jpg" in the data directory corresponds with "10.png" in the ground truth directory and with index 10 in the json files (= label_data_all.json and Curve_parameters.json).
You can download the complete TuSimple dataset from Excerpt of 7,821 characters
Read on GitHub
Wouter Van Gansbeke · Google DeepMind · United Kingdom
37
Would you bet a product on this? Bounded 0–100 and slow moving.
matched fp:c3d8881a1187ad80, topic:deep-learning, topic:pytorch
matched fp:c3d8881a1187ad80, topic:computer-vision, readme:object detection