RAIS Robust and Accurate Interactive Segmentation via Continual Learning Yuying Hao Yi Liu Juncai Peng Haoyi Xiong Guowei Chen

2025-04-29 0 0 2.15MB 8 页 10玖币

侵权投诉

RAIS: Robust and Accurate Interactive Segmentation

via Continual Learning

Yuying Hao, Yi Liu, Juncai Peng, Haoyi Xiong, Guowei Chen,

Shiyu Tang, Zeyu Chen, Baohua Lai

Baidu Inc.

{haoyuying, liuyi22}@baidu.com

Abstract

Interactive image segmentation aims at segmenting a

target region through a way of human-computer interaction.

Recent works based on deep learning have achieved excel-

lent performance, while most of them focus on improving

the accuracy of the training set and ignore potential im-

provement on the test set. In the inference phase, they tend

to have a good performance on similar domains to the train-

ing set, and lack adaptability to domain shift, so they require

more user efforts to obtain satisfactory results. In this work,

we propose RAIS, a robust and accurate architecture for in-

teractive segmentation with continuous learning, where the

model can learn from both train and test data sets. For efﬁ-

cient learning on the test set, we propose a novel optimiza-

tion strategy to update global and local parameters with a

basic segmentation module and adaptation module, respec-

tively. Moreover, we perform extensive experiments on sev-

eral benchmarks that show our method can handle data dis-

tribution shifts and achieves SOTA performance compared

with recent interactive segmentation methods. Besides, our

method also shows its robustness in the datasets of remote

sensing and medical imaging where the data domains are

completely different between training and testing.

1. Introduction

Deep learning methods have shown superior perfor-

mance on segmentation tasks [32, 13, 18], such as por-

trait segmentation [3], satellite image processing [10], in-

telligent driving [29]. Unusually, most of them require

large-scale annotated images to learn powerful abstraction.

However, the cost of manual annotation grows rapidly, as

the number of data increases, especially when it comes

to pixel-level segmentation tasks. To improve the efﬁ-

ciency of the annotation process, interactive segmentation

appears to be an effective and auxiliary way, which is a

semi-automatic method utilizing human-computer interac-

tion. It allows the annotators to provide a small number

of interactive information and generates the ﬁnal segmenta-

tion result progressively. Therefore, it can accelerate seg-

mentation annotation while maintaining satisfactory qual-

ity. Recently, interactive segmentation has attracted inten-

sive attention in both academia and industry. In interac-

tive segmentation, there have been a few types of inter-

active information, e.g., bounding box [30], scribbles [1]

or clicks [27, 22, 11, 17, 26], where their characteristics

have been studied well by previous works. Among them,

the click-based interactive way is the most widely used,

because it provides sufﬁcient region-of-interest informa-

tion with minimal interaction time. In general, click-based

methods usually employ two kinds of user clicks, i.e. posi-

tive clicks and negative clicks, which indicate the target re-

gion and non-target regions, respectively. In general, most

interactive methods [27, 22, 17] train the model over a train-

ing set without updating its parameters at test time. Usually,

they do well on the test data similar to the train set. As the

difference in data distribution increases, their performance

could deteriorate signiﬁcantly. Accordingly, they require

more user clicks to reﬁne the ﬁnal results, or even they need

to be re-trained on the new data, which is increasing anno-

tation costs.

In this work, we propose RAIS, a robust and accu-

rate architecture for interactive segmentation with contin-

uous learning, to address the deterioration problem. In our

method, we take interactive segmentation as a continuous

adaptation and allow the model to learn from both the train

set and test set. For the train set, we use the full-supervision

way to update the model parameters like other methods. As

for the test set, we propose a weakly-supervised method to

reﬁne the model by utilizing the user annotations and in-

termediate output. Since the user interactions have already

provided useful hints of ground truth, the intermediate re-

sults are the potential to improve performance for subse-

quent data. Hence, our model can adapt to the new data

distribution gradually, and relieve the impact of the deteri-

oration problem. Also, to prevent the model from forget-

arXiv:2210.10984v1 [cs.CV] 20 Oct 2022

Figure 1. Overview of the RAIS architecture. It consists of two parts: a basic segmentation module (BSM) and an adaptation module

(ADM).

ting previous knowledge when learning new information,

we need to minimize unnecessary changes to model param-

eters at the test time. Therefore, we propose a novel opti-

mization strategy for updating two parts of the model, i.e.

basic segmentation module (BSM) and the adaptation mod-

ule (ADM). BSM retains fundamental knowledge learning

from the initial train set, and it would not be updated fre-

quently. ADM learns the residual feature representation be-

tween the train set and the test set, and it is updated more

frequently to ﬁt new data distribution. Compared to opti-

mizing the overall model directly, our strategy improves the

model robustness. With the proposed continuous adaptation

strategy, the comprehensive evaluations show our superior

performance on well-known benchmarks.

Our contributions are summarised as follows:

• We propose a robust and accurate architecture for

interactive segmentation with continual learning.

Through continuous adaptation, the model ﬁts the new

data distribution gradually and relieves the deteriora-

tion problem of distribution shift.

• We propose a novel optimization strategy for interac-

tive segmentation tasks at test time. By minimizing un-

necessary changes to model parameters, the model pre-

vents from forgetting previous knowledge when learn-

ing new information.

• Extensive experiments demonstrate that our method

achieves SOTA performance on several benchmarks.

Moreover, our method can improve the robustness of

domain changes with fewer user clicks.

2. Related Works

The interactive segmentation task aims to obtain an ac-

curate mask of an object with minimal user interaction. In-

teractive information can be clicks [26, 11, 27], bounding

boxes [30], extreme points [22] or scribbles [1]. Interac-

tive image segmentation has existed for decades. Tradi-

tional methods usually optimize energy functions to obtain

the object of interest. These methods utilize only low-level

features to distinguish foreground and background, so they

are not either accurate or robust.

Recently, interactive segmentation based on deep learn-

ing has developed rapidly. According to whether the model

changes its parameters dynamically on user clicks and test

data, these methods can be roughly divided into two types:

off-the-shelf methods and on-the-ﬂy methods.

Off-the-shelf methods optimize the model parameters

in the training process and freeze parameters when testing.

Xu et al. [31] took clicks as additional input to ﬁne-tune the

fcn [19] network. It is the ﬁrst work introducing deep learn-

ing into the interactive image segmentation task. Maninis

et al. [22] adopted extreme points to extract the region of

interest, they considered that annotating extreme points can

reduce annotation time and improve the segmentation ac-

curacy compared with bounding box annotations. Chen et

al. [2] reﬁned the segmentation result by conditional dif-

fusing information of user clicks. They reﬁned the self-

attention strategy by adding click and mask information to

the afﬁnity matrix and making the model understand user

intentions easily. Soﬁiuk et al. [27] found that a training set

has a great impact on interactive performance. They pro-

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：

RAIS:RobustandAccurateInteractiveSegmentationviaContinualLearningYuyingHao,YiLiu,JuncaiPeng,HaoyiXiong,GuoweiChen,ShiyuTang,ZeyuChen,BaohuaLaiBaiduInc.{haoyuying,liuyi22}@baidu.comAbstractInteractiveimagesegmentationaimsatsegmentingatargetregionthroughawayofhuman-computerinteraction.Recentworksbased...

展开>> 收起<<

RAIS Robust and Accurate Interactive Segmentation via Continual Learning Yuying Hao Yi Liu Juncai Peng Haoyi Xiong Guowei Chen.pdf

共8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！

RAIS Robust and Accurate Interactive Segmentation via Continual Learning Yuying Hao Yi Liu Juncai Peng Haoyi Xiong Guowei Chen

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: