Pointly-Supervised Panoptic Segmentation Junsong Fan13 Zhaoxiang Zhang123 and Tieniu Tan12 1Center for Research on Intelligent Perception and Computing

2025-05-02 0 0 4.21MB 18 页 10玖币

侵权投诉

Pointly-Supervised Panoptic Segmentation

Junsong Fan1,3, Zhaoxiang Zhang*1,2,3, and Tieniu Tan1,2

1Center for Research on Intelligent Perception and Computing,

Institute of Automation, Chinese Academy of Sciences, Beijing, China

2University of Chinese Academy of Sciences, Beijing, China

3Centre for Artiﬁcial Intelligence and Robotics, HKISI CAS, HongKong, China

{fanjunsong2016@,zhaoxiang.zhang@,tnt@nlpr.}ia.ac.cn

Abstract. In this paper, we propose a new approach to applying point-

level annotations for weakly-supervised panoptic segmentation. Instead

of the dense pixel-level labels used by fully supervised methods, point-

level labels only provide a single point for each target as supervision, sig-

niﬁcantly reducing the annotation burden. We formulate the problem in

an end-to-end framework by simultaneously generating panoptic pseudo-

masks from point-level labels and learning from them. To tackle the core

challenge, i.e., panoptic pseudo-mask generation, we propose a principled

approach to parsing pixels by minimizing pixel-to-point traversing costs,

which model semantic similarity, low-level texture cues, and high-level

manifold knowledge to discriminate panoptic targets. We conduct exper-

iments on the Pascal VOC and the MS COCO datasets to demonstrate

the approach’s eﬀectiveness and show state-of-the-art performance in the

weakly-supervised panoptic segmentation problem. Codes are available

at https://github.com/BraveGroup/PSPS.git.

Keywords: weakly-supervised learning, panoptic segmentation

1 Introduction

Panoptic segmentation [23] aims at fully parsing all the pixels into nonoverlap-

ping masks for both thing instances and stuﬀ classes. It combines the semantic

segmentation and the instance segmentation tasks simultaneously. Classical deep

learning approaches require precise dense pixel-level labels to solve this problem.

However, acquiring exact pixel- and instance-level annotations on large-scale

datasets is very time-consuming, hindering the popularization and generaliza-

tion of the approaches in new practical applications.

To alleviate the annotation burden for segmentation models, researchers re-

cently proposed weakly-supervised learning (WSL) [4,52,51], which focuses on

leveraging coarse labels to train dense pixel-level segmentation tasks. Typically,

the weak supervision includes image-level [14,16,15], point-level [2,38], scribble-

level [47,31], and bounding box-level labels [9], etc. These approaches tackle

either semantic segmentation [36], instance segmentation [1,21], or panoptic seg-

mentation [41,27] tasks. Among them, the weakly-supervised panoptic segmen-

tation (WSPS) problem is the most challenging since it requires both semantic

arXiv:2210.13950v1 [cs.CV] 25 Oct 2022

2 J. Fan et al.

and instance discrimination with only weak supervision. As a result, the WSPS

got less attention in previous works, and its performance is far from satisfactory.

The seminal work by Li et al. [27] manages to address the WSPS problem using

bounding-box level labels. Later, JTSM [41] proposes to apply only image-level

labels for the WSPS problem. Recently, PanopticFCN [29] tackles this problem

by connecting multiple point labels into polygons. The performances of these

approaches diﬀer signiﬁcantly with the diﬀerent weak annotations.

In this paper, we propose a new WSPS paradigm to use only a single point

for each target as the supervision, as illustrated in Fig. 1. Recall that the core of

weakly-supervised segmentation is to release the annotation burden while still

obtaining decent performance. In other words, balance the cost of annotation and

the model performance. We are motivated to use the point-level labels because,

on the one hand, the annotation time of point-level labels is only marginally

above the image-level labels [2], saving much cost compared with the box-level

or polygon labels. On the other hand, point labels can provide minimum spa-

tial information to localize and discriminate diﬀerent panoptic targets for the

segmentation models.

A natural idea to estimate panoptic masks from point-level labels is to as-

sign each pixel in the image to one of the points according to some principles.

To this end, we propose tackling this problem by minimizing the pixel-to-point

traversing cost, measured by the neighboring pixel aﬃnities. There are two basic

requirements to correctly assign pixels to point labels: semantic class matching

and instance discrimination. The former ensures that the pixels are parsed with

the correct class labels, and the latter is responsible for distinguishing diﬀerent

instances in the thing classes. Therefore, we consider three criteria to model the

aﬃnities: semantic similarity, low-level image cues, and high-level instance dis-

crimination knowledge. Using these criteria, we model the pixel-to-point travers-

ing costs and solve the assignment problem by ﬁnding the shortest path.

We base our approach on the transformer models [46,11,39], which have re-

cently shown impressive results in computer vision tasks [8,30,44,3,56]. Speciﬁ-

cally, our approach contains a group of semantic query tokens to parse semantic

segmentation results and a group of panoptic query tokens responsible for the

panoptic segmentation task [30]. In addition to the regular panoptic segmen-

tation model, our approach contains a label generation model, which produces

dense panoptic pseudo-masks depending on the point-level labels and the crite-

ria above. The whole approach is end-to-end. After training, only the panoptic

segmentation branch is kept for testing. Thus, it does not incur additional com-

putation or memory overhead for usage. We conduct thorough experiments to

analyze the proposed approach and the properties of the point-level labels. Mean-

while, we demonstrate new cutting-edge performance with the WSPS problem

on the Pascal VOC [13] and the MS COCO [32] datasets.

In summary, the main contributions of this work are:

–We propose a new paradigm for the WSPS problem, which utilizes a single

point for each target as supervision for training.

Pointly-Supervised Panoptic Segmentation 3

image point label panoptic segmentation

Fig. 1. Illustration of the proposed pointly-supervised panoptic segmentation. From

left to right: input images, point labels as supervision, and panoptic segmentation

predictions. The point labels provide a single point annotation for each target, including

both thing instances and stuﬀ classes, which are used at training time only. Please see

Sec. 3for details. Best viewed in color.

–A novel approach to estimating dense panoptic pseudo-masks by minimizing

the pixel-to-point traversing distance is proposed.

–We implement the approach in an end-to-end framework with transformers,

conduct analytical experiments to study the model and the point-level labels,

and demonstrate state-of-the-art performance on the Pascal VOC and the

MS COCO datasets.

2 Related Works

2.1 Panoptic Segmentation

The panoptic segmentation [23] task simultaneously incorporates semantic seg-

mentation and instance segmentation, where each pixel is uniquely assigned

to one of the stuﬀ classes or one of the thing instances. This problem can

be tackled by combining the semantic and instance segmentation results in a

post-processing manner [23]. Later works such as JSIS [10] adopt a uniﬁed net-

work combining a semantic segmentation branch and an instance segmentation

branch. After that, many approaches have been proposed for improvement by

using feature pyramids [22], automatic architecture searching [49], and unifying

the pipeline [28], etc.

Recently, transformer-based approaches have shown impressive results across

the NLP [46,11,39] and the CV [3,12,56,34] applications. The seminal work

DETR [3] provides a clear and elegant solution for object detection and seg-

mentation. The following work DeformableDETR [56] improves it by using the

deformable transformers to reduce the computation burden and accelerate the

convergence. K-Net [54] adopts an iterative reﬁnement procedure to enhance

the attention masks gradually. MaskFormer [8] proposes to separate the mask

prediction and the classiﬁcation process. Panoptic SegFormer [30] embraces a

similar idea and adopts an auxiliary localization target to ease the model train-

ing. Our panoptic segmentation approach is based on these works, and we focus

on alleviating the annotation burden by exploiting point-level annotations.

4 J. Fan et al.

2.2 Weakly-Supervised Segmentation

Weakly supervised segmentation [4,52] aims to alleviate the annotation bur-

den for segmentation tasks by using weak labels for training. According to the

type of tasks, it concerns semantic segmentation [16,36,24,48,26,17], instance

segmentation [1,45], and panoptic segmentation [27,41] problems. According to

the kinds of supervision, these approaches use image-level [16,36,24,48,26,41],

point-level [2,38], scribble-level [47,31], or box-level [43,45,9] labels for training.

Among them, image-level label-based approaches are the most prevalent. These

approaches generally rely on the CAM [55,40] to extract spatial information from

classiﬁcation models, which are trained by the image-level labels. Though great

progress has been achieved by these approaches on the semantic segmentation

task, it is generally hard to distinguish diﬀerent instances of the same class with

only image-level labels, especially on large-scale datasets with many overlapping

instances. Li et al. [27] proposes to address this problem by additionally using

bounding-box annotations, which however takes much more time to annotate.

PanopticFCN [29] alternatively proposes to use coarse polygons to supervise the

panoptic segmentation model, which are obtained by connecting multiple point

annotations for each target. PSIS [7] proposes to address the instance segmen-

tation problem by using sparsely sampled foreground and background points in

each bounding box. Though these approaches achieve better results, their an-

notation burden is signiﬁcantly heavier than image-level labels. In this paper,

we try to use a new form of weak annotation for panoptic segmentation, i.e., a

single point for each target. We demonstrate that this supervision can achieve

competitive performance compared with previous approaches while signiﬁcantly

reducing the annotation burden.

2.3 Point-Level Labels in Visual Tasks

Recently point-level annotation has drawn interest in a broad range of com-

puter vision tasks. Beside the works concerning the detection and segmenta-

tion tasks [7,29,38,2], some works adopt point-level labels to train crowd count-

ing [50,33] models. SPTS [37] proposes to use points for the text spotting prob-

lem. Chen et al. [5] propose addressing weakly-supervised detection problems

using point labels. Besides, point labels also play an essential role in interactive

segmentation models, where users provide interactive hints through point-level

clicks [53,35,42]. To the best of our knowledge, there are still no approaches to

training panoptic segmentation models using only a single point per target.

3 Approach

In this section, we elaborate on the details of the proposed approach. Fig. 2

illustrates the overall framework, which can be decomposed into two major com-

ponents, a label generation model and a panoptic segmentation model. These

two components share the same backbone and the transformer encoder [56]. The

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：

Pointly-SupervisedPanopticSegmentationJunsongFan1,3,ZhaoxiangZhang*1,2,3,andTieniuTan1,21CenterforResearchonIntelligentPerceptionandComputing,InstituteofAutomation,ChineseAcademyofSciences,Beijing,China2UniversityofChineseAcademyofSciences,Beijing,China3CentreforArtificialIntelligenceandRobotics,HKI...

展开>> 收起<<

Pointly-Supervised Panoptic Segmentation Junsong Fan13 Zhaoxiang Zhang123 and Tieniu Tan12 1Center for Research on Intelligent Perception and Computing.pdf

共18页,预览4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！

Pointly-Supervised Panoptic Segmentation Junsong Fan13 Zhaoxiang Zhang123 and Tieniu Tan12 1Center for Research on Intelligent Perception and Computing

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: