目标检测:Proposal-Contrastive Pretraining for Object Detection from Fewer Data

目标检测:Proposal-Contrastive Pretraining for Object Detection from Fewer Data_第1张图片

论文作者:Quentin Bouniot,Romaric Audigier,Angélique Loesch,Amaury Habrard

作者单位:Université Paris-Saclay; Université Jean Monnet Saint-Etienne; Universitaire de France (IUF)

论文链接:http://arxiv.org/abs/2310.16835v1

内容简介:

1)方向:目标检测

2)应用:目标检测

3)背景:在目标检测中,使用预训练的深度神经网络是一种有效的方法,但对于无监督预训练,通常使用大批量数据来进行对比学习,需要大量资源。

4)方法:本文提出 ProSeCo,一种新的无监督预训练方法。该方法利用目标检测器生成的大量目标建议进行对比学习,这允许使用较小的批量大小,并结合目标级特征来学习图像中的局部信息。为了改善对比损失的效果,研究引入了对象位置信息,以考虑多个重叠的目标建议。此外,研究还强调了在重用预训练骨干网络时,需要保持骨干网络和检测头之间的局部信息的一致性。

5)结果:结果表明,ProSeCo方法在标准和新的基准数据集上,比当前领先的无监督目标检测预训练方法表现更出色,尤其是在有限数据的情况下学习方面。

目标检测:Proposal-Contrastive Pretraining for Object Detection from Fewer Data_第2张图片

目标检测:Proposal-Contrastive Pretraining for Object Detection from Fewer Data_第3张图片

目标检测:Proposal-Contrastive Pretraining for Object Detection from Fewer Data_第4张图片

目标检测:Proposal-Contrastive Pretraining for Object Detection from Fewer Data_第5张图片

目标检测:Proposal-Contrastive Pretraining for Object Detection from Fewer Data_第6张图片

目标检测:Proposal-Contrastive Pretraining for Object Detection from Fewer Data_第7张图片

目标检测:Proposal-Contrastive Pretraining for Object Detection from Fewer Data_第8张图片

目标检测:Proposal-Contrastive Pretraining for Object Detection from Fewer Data_第9张图片

目标检测:Proposal-Contrastive Pretraining for Object Detection from Fewer Data_第10张图片

目标检测:Proposal-Contrastive Pretraining for Object Detection from Fewer Data_第11张图片

目标检测:Proposal-Contrastive Pretraining for Object Detection from Fewer Data_第12张图片

目标检测:Proposal-Contrastive Pretraining for Object Detection from Fewer Data_第13张图片

目标检测:Proposal-Contrastive Pretraining for Object Detection from Fewer Data_第14张图片

目标检测:Proposal-Contrastive Pretraining for Object Detection from Fewer Data_第15张图片

你可能感兴趣的:(目标检测,目标跟踪,人工智能)