NON-ITERATIVE OPTIMIZATION OF PSEUDO-LABELING THRESHOLDS FOR TRAINING OBJECT DETECTION MODELS FROM MULTIPLE DATASETS Yuki Tanaka Shuhei M. Yoshida Makoto Terao

2025-05-02 0 0 3.09MB 5 页 10玖币

侵权投诉

NON-ITERATIVE OPTIMIZATION OF PSEUDO-LABELING THRESHOLDS FOR

TRAINING OBJECT DETECTION MODELS FROM MULTIPLE DATASETS

Yuki Tanaka, Shuhei M. Yoshida, Makoto Terao

Visual Intelligence Research Laboratories

NEC Corporation

Kawasaki, Kanagawa, Japan

ABSTRACT

We propose a non-iterative method to optimize pseudo-

labeling thresholds for learning object detection from a col-

lection of low-cost datasets, each of which is annotated for

only a subset of all the object classes. A popular approach

to this problem is ﬁrst to train teacher models and then to

use their conﬁdent predictions as pseudo ground-truth labels

when training a student model. To obtain the best result,

however, thresholds for prediction conﬁdence must be ad-

justed. This process typically involves iterative search and

repeated training of student models and is time-consuming.

Therefore, we develop a method to optimize the thresholds

without iterative optimization by maximizing the Fβ-score

on a validation dataset, which measures the quality of pseudo

labels and can be measured without training a student model.

We experimentally demonstrate that our proposed method

achieves an mAP comparable to that of grid search on the

COCO and VOC datasets.

Index Terms—Non-iterative optimization, pseudo label-

ing, object detection, weakly supervised learning

1. INTRODUCTION

Object detection [1, 2, 3, 4] has achieved signiﬁcant progress

in deep learning with a tremendous number of images and an-

notations, but it becomes quite expensive to collect them. This

creates a signiﬁcant barrier when it comes to moving from the

research stage to practical application. Recently, research on

how to train a model with low-cost datasets has become more

active.

There are several paradigms to learn from low-cost

datasets. Examples include semi-supervised learning [5, 6, 7,

8] and weakly supervised learning [9, 10]. In semi-supervised

learning, models are trained from a limited amount of labeled

data and a lot of unlabeled data (Fig. 1(a)), while in weakly

from IEEE must be obtained for all other uses, in any current or future media,

including reprinting/republishing this material for advertising or promotional

purposes, creating new collective works, for resale or redistribution to servers

or lists, or reuse of any copyrighted component of this work in other works.

(a) Semi-supervised learning (b) Weakly supervised learning

Fig. 1. Examples of learning paradigms for object detection

using low-cost datasets. In these examples, the goal is to

train a model that detects people and bicycles in images by

using training datasets with annotations as illustrated above.

The two images are contained in COCO [14] and VOC [15]

datasets, respectively.

supervised learning, models are trained from only image-

level annotations and no bounding boxes (Fig. 1(b)). By

contrast, we aim at training a single object detection model

for all classes from multiple datasets that have different class

sets without additional annotations [11, 12, 13]. This set-

ting (Fig. 1(c)) is important for practical applications, because

we can add object categories simply by combining datasets

that are made for different purposes.

Typically, pseudo labeling is used to train an object de-

tection model in the current problem setting (Fig. 2). Specif-

ically, we ﬁrst train one teacher model from each dataset and

then use them to predict locations of unlabeled objects. A

prediction is used as a pseudo label if its conﬁdence score

is higher than a predetermined threshold. Finally, we train a

single student model for all classes by using both the ground-

truth labels and the pseudo labels.

To achieve the best performance with pseudo labeling, it is

imperative to decide this threshold properly, but the optimiza-

arXiv:2210.10221v1 [cs.CV] 19 Oct 2022

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：

NON-ITERATIVEOPTIMIZATIONOFPSEUDO-LABELINGTHRESHOLDSFORTRAININGOBJECTDETECTIONMODELSFROMMULTIPLEDATASETSYukiTanaka,ShuheiM.Yoshida,MakotoTeraoVisualIntelligenceResearchLaboratoriesNECCorporationKawasaki,Kanagawa,JapanABSTRACTWeproposeanon-iterativemethodtooptimizepseudo-labelingthresholdsforlearning...

展开>> 收起<<

NON-ITERATIVE OPTIMIZATION OF PSEUDO-LABELING THRESHOLDS FOR TRAINING OBJECT DETECTION MODELS FROM MULTIPLE DATASETS Yuki Tanaka Shuhei M. Yoshida Makoto Terao.pdf

共5页,预览1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！

NON-ITERATIVE OPTIMIZATION OF PSEUDO-LABELING THRESHOLDS FOR TRAINING OBJECT DETECTION MODELS FROM MULTIPLE DATASETS Yuki Tanaka Shuhei M. Yoshida Makoto Terao

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: