Repainting and Imitating Learning for Lane Detection Yue He heyue04baidu.com

2025-04-29 0 0 4.52MB 9 页 10玖币

侵权投诉

Repainting and Imitating Learning for Lane Detection

Yue He∗

heyue04@baidu.com

Baidu Inc.

Beijing, China

Minyue Jiang∗

jiangminyue@baidu.com

Baidu Inc.

Beijing, China

Xiaoqing Ye∗

yexiaoqing@baidu.com

Baidu Inc.

Shanghai, China

Liang Du∗

duliang@mail.ustc.edu.cn

Fudan University

Shanghai, China

Zhikang Zou

zouzhikang@baidu.com

Baidu Inc.

Shenzhen, China

Wei Zhang

zhangwei99@baidu.com

Baidu Inc.

Shenzhen, China

Xiao Tan

tanxiao01@baidu.com

Baidu Inc.

Shenzhen, China

Errui Ding

dingerrui@baidu.com

Baidu Inc.

Beijing, China

ABSTRACT

Current lane detection methods are struggling with the invisibility

lane issue caused by heavy shadows, severe road mark degradation,

and serious vehicle occlusion. As a result, discriminative lane fea-

tures can be barely learned by the network despite elaborate designs

due to the inherent invisibility of lanes in the wild. In this paper, we

target at nding an enhanced feature space where the lane features

are distinctive while maintaining a similar distribution of lanes in

the wild. To achieve this, we propose a novel Repainting and Imi-

tating Learning (RIL) framework containing a pair of teacher and

student without any extra data or extra laborious labeling. Speci-

cally, in the repainting step, an enhanced ideal virtual lane dataset

is built in which only the lane regions are repainted while non-lane

regions are kept unchanged, maintaining the similar distribution

of lanes in the wild. The teacher model learns enhanced discrimi-

native representation based on the virtual data and serves as the

guidance for a student model to imitate. In the imitating learning

step, through the scale-fusing distillation module, the student net-

work is encouraged to generate features that mimic the teacher

model both on the same scale and cross scales. Furthermore, the

coupled adversarial module builds the bridge to connect not only

teacher and student models but also virtual and real data, adjusting

the imitating learning process dynamically. Note that our method

introduces no extra time cost during inference and can be plug-and-

play in various cutting-edge lane detection networks. Experimental

results prove the eectiveness of the RIL framework both on CU-

Lane and TuSimple for four modern lane detection methods. The

code and model will be available soon.

∗Equal contribution.

Permission to make digital or hard copies of all or part of this work for personal or

classroom use is granted without fee provided that copies are not made or distributed

for prot or commercial advantage and that copies bear this notice and the full citation

on the rst page. Copyrights for components of this work owned by others than ACM

must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,

to post on servers or to redistribute to lists, requires prior specic permission and/or a

fee. Request permissions from permissions@acm.org.

MM ’22, October 10–14, 2022, Lisboa, Portugal

ACM ISBN 978-1-4503-9203-7/22/10. . . $15.00

https://doi.org/10.1145/3503161.3548042

CCS CONCEPTS

•Computing methodologies

Interest point and salient

region detections;

KEYWORDS

lane detection, image repainting, scale-fusing distillation, coupled

adversarial module

ACM Reference Format:

Yue He, Minyue Jiang, Xiaoqing Ye, Liang Du, Zhikang Zou, Wei Zhang,

Xiao Tan, and Errui Ding. 2022. Repainting and Imitating Learning for

Lane Detection. In Proceedings of the 30th ACM International Conference on

Multimedia (MM ’22), October 10–14, 2022, Lisboa, Portugal. ACM, New York,

NY, USA, 9 pages. https://doi.org/10.1145/3503161.3548042

1 INTRODUCTION

Lane detection is a crucial task in autonomous driving [

], which

could serve as visual cues for advanced driver assistance systems

(ADAS) to keep vehicles stably following lane markings. Thus,

autonomous vehicles need to locate each lane’s precise position.

With the development of deep learning, neural networks [

]

have been widely used in lane detection for their compelling per-

formance. Early deep-learning-based methods detect lanes through

pixel-wise segmentation based framework [

], where each

pixel is assigned a binary label to indicate whether it belongs to a

lane or not. More recently, various anchor-based methods [

]

are proposed, dierent forms of anchors such as line anchors and

box anchors are used among these methods, aiming to let the

networks focus the optimization on the line shape by regress-

ing the relative coordinates. Besides, row-wise classication meth-

ods [

] rely on the shape prior of lanes and predict

the location for each row. Parametric prediction methods [

]

directly output parameters of curve equation for lines. Multi-task

learning is further combined to improve lane detection accuracy in

a complicated environment. For example, VPGNet [

] integrates

road marking detection and vanishing point prediction to obtain

auxiliary information for lane detection. However, additional man-

ual labeling is time-consuming and laborious.

MM ’22, October 10–14, 2022, Lisboa, Portugal Yue He et al.

(a)

(b)

(c)

(d)

Figure 1: Samples of complex scenarios in the CULane dataset:

(a) Crowded (b) Dazzle (c) Shadow (d) Night. The �rst column

shows the ground truth, and we highlight the elliptical re-

gion for comparison between lines before repainting (second

column) and after repainting (third column).

Either combining shape prior of lanes or designing the auxiliary

task has shown its comprehensive consideration and competitive

results for lane detection, it is still a challenging task due to many

factors such as the wide variety of lane markings appearance includ-

ing solid or dashed, white or yellow. Moreover, complex road and

light conditions leading to occlusion and low illumination increase

the invisibility of lanes. All these diculties require the method to

have the ability to extract lanes under complicated environments.

Taking CULane dataset [

] as an example in Fig. 1, we present

four representative scenarios, including crowded, dazzle, shadow,

and night. In these scenarios, complete lanes cannot be visually

detected by comparing lanes in the wild (the second column) with

the ground truth annotations (the rst column). The inherent in-

visibility of lanes hinders the progress of the algorithms. In this

paper, we focus on nding a more discriminative lane feature space

and maintaining a similar distribution of lanes in the wild. The

Repainting and Imitating Learning (RIL) framework is proposed to

increase the visibility of lanes by repainting module without extra

data or labor labeling, and improve feature discrimination while

transferring lane knowledge from teacher to student via imitating

learning simultaneously.

First and foremost, the inherent invisibility of lanes in the wild

intractable by current lane detection methods is alleviated by a sim-

ple but eective repainting module in our repainting step. Through

this module, virtual data is generated based on the location of lanes

annotated in the ground truth without the requirement of extra data

or laborious labeling. This module highlights the fuzzy lanes and

makes lines more prominent and continuous. As shown in Fig. 1,

the lane regions in the third column become more distinctive and

continuous while maintaining other non-lane regions unchanged.

Based on these ideal lanes, the teacher will be trained in advanced

and achieves upper bound performance.

A simple solution by directly adding the virtual data as data

augmentation does not work, since there exists a distribution gap

between virtual data and real data. Thus, to better utilize virtual data,

imitating learning step is introduced including a scale-fusing distil-

lation module and a coupled adversarial module. In the scale-fusing

distillation module, dierent stages of feature representations from

the teacher are treated as the ideal enhanced feature space. Noticing

that the teacher model has the same architecture as the student.

The teacher’s feature maps of the same size as the student’s feature

maps are distilled directly. Besides, the teacher’s larger feature maps

are down-sampled to distill student’s semantic feature maps simul-

taneously for ner lane details. Both the same scale and cross scale

information are distilled, helping student to imitate the teacher.

To further eliminate distribution gaps not only between dierent

networks but also between dierent input data, coupled adversarial

module is proposed to build a bridge to connect networks as well

as data. A pair of discriminators are coupled by adding in another

student’s output of virtual data. The rst net-sensitive discrimina-

tor is to distinguish teacher and student networks when they are

fed into virtual data. The second data-sensitive discriminator is to

distinguish between virtual data and real data which feed to the

student network. By coupling two discriminators, the student can

better imitate the enhanced teacher features dynamically through

the learning process.

Our main contributions are summarized as follows:

•

We introduce a simple yet eective Repainting and Imitating

framework (RIL) for lane detection, focusing on discriminat-

ing lane features and maintaining the similar distribution of

lanes in the wild by nding an enhanced feature space.

•

We repaint the real lane data to ideal virtual data in the

repainting step, achieving enhanced representation under

complicated environments.

•

We combine the scale-fusing distillation module with the

coupled adversarial module in the imitating step, building the

bridge between networks and data to weaken the learning

gap.

The proposed RIL framework can be easily plug-and-play in most

cutting-edge methods without any extra inference cost. Experi-

mental results prove the eectiveness of RIL framework both on

CULane [

] and TuSimple [

] for four modern lane detection

methods including UFAST [

], ERFNet [

], ESA [

] and Cond-

LaneNet [14] respectively.

2 RELATED WORK

Lane detection Recent approaches [

] focus on the deep

neural networks and signicantly boost the lane detection perfor-

mance due to the powerful representation learning ability. Some

methods[

] treat lane detection as a semantic segmentation

task. For instance, SCNN [

] designs slice-by-slice convolutions

within feature maps to exchange pixel information between pixels

across rows and columns in a layer. Inspired by network architec-

ture search (NAS), CurveLanes-NAS [

] designs a lane-sensitive

architecture to incorporate both long-ranged coherent lane infor-

mation and short-ranged local lane information. Despite the promis-

ing results, the computational complexity in these methods brings

heavy inference overhead. Therefore, row-wise classication based

methods [

] have been proposed for ecient lane detec-

tion. These approaches divide the input image into grids and predict

the most probable cell to contain a part of a lane, which realize the

trade-obetween speed and accuracy. More recently, SGNet [

]

introduces a structure-guided framework to accurately classify, lo-

cate and restore the shape of unlimited lanes. Unlike the systematic

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

RepaintingandImitatingLearningforLaneDetectionYueHe∗heyue04@baidu.comBaiduInc.Beijing,ChinaMinyueJiang∗jiangminyue@baidu.comBaiduInc.Beijing,ChinaXiaoqingYe∗yexiaoqing@baidu.comBaiduInc.Shanghai,ChinaLiangDu∗duliang@mail.ustc.edu.cnFudanUniversityShanghai,ChinaZhikangZouzouzhikang@baidu.comBaiduInc....

展开>> 收起<<

Repainting and Imitating Learning for Lane Detection Yue He heyue04baidu.com.pdf

共9页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Repainting and Imitating Learning for Lane Detection Yue He heyue04baidu.com

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: