1.College of Engineering and Design,Hunan Normal University,Changsha 410081,Hunan,China
2.The Key Laboratory of Intelligent Sensing and Rehabilitation Robotics of Hunan Province Universities,Changsha 410081,Hunan,China
3.Jiangxi Communication Terminal Industrial Technology Research Institute Co. Ltd. ,Ji’an 343600,Jiangxi,China
Show less
文章历史+
Received
Published
2024-05-23
2025-10-24
Issue Date
2026-07-23
PDF (3304K)
摘要
开放世界目标检测(Open World Object Detection, OWOD)旨在检测环境中的未见目标,是智慧教育领域中诸多问题的上游任务。现有的方法通过选择前K个置信度最高的前景进行预测,直接将在已知类别上训练的检测器泛化到检测其他的未知前景目标上。然而这种方法使检测器偏向检测已知类别,难以识别全部未知目标。对此,提出了一种基于平滑性无监督领域自适应的开放世界目标检测方法。首先,利用两阶段目标检测Faster-RCNN为每一张图像生成候选区域,将目标得分高与目标得分极低的候选区域作为源域,将剩余候选区域作为目标域。然后,采用平滑性无监督领域自适应方法,学习前后景分类器,该方法通过引入平滑域对抗训练过程,能有效地提高域对抗方法在分类任务中的性能,实现对已知和未知类别的前景目标无偏向预测。在MS-COCO数据集以及自行构建的化学实验器材检测数据集上进行模型性能的评估实验。结果表明,相比于直接选择置信度最高的预测,本文提出的方法能更好地平衡已知和未知目标类别的检测,提高检测器的整体性能。
Abstract
Open World Object Detection (OWOD) aims to detect unseen objects in an environment and serves as an upstream task for various issues in intelligent education. Existing methods predict the top K most confident foregrounds by generalizing the detector, which was trained on known categories, to detect other unknown foreground objects. However, this approach tends to bias the detector towards known categories, hindering the identification of all unknown objects. To address this limitation, we propose an open world object detection method based on Smoothness Unsupervised Domain Adaptation. First, we use a two-stage object detector, Faster-RCNN, to generate candidate regions for each image. The regions with the highest and lowest object scores are considered the source domain, while the remaining regions are treated as the target domain. Then, we employ a smoothness unsupervised domain adaptation method to train a foreground-background classifier. This method introduces a smooth domain adversarial training process, enhancing the performance of domain adversarial methods in classification tasks. By utilizing this approach, we achieve unbiased predictions of foreground objects for both known and unknown categories. This article evaluates the performance of the model on the MS-COCO dataset and a self constructed chemical experimental equipment detection dataset. Experiments demonstrate that, compared to directly selecting the highest confidence predictions, the proposed method better balances the detection of known and unknown objects, improving overall detector performance.
检测未知物体是将环境中的未见目标检测出来,并将其类别标记为“未知(unknown)”,它是开放世界目标检测(Open World Object Detection, OWOD)的关键步骤,也是后续步骤的基础[2]。这一问题的难点是缺乏对未知类别的标签作为有效监督信息。现有方法通常采用Top-K选择策略生成标签。然而,Top-K选择策略存在两个缺点:首先,超参数K的选择需要通过手动调整实现,缺乏自适应性;其次,目标得分是使用已知类别训练的检测器计算出来的,计算结果不可避免地会偏向已知类别,同时倾向于为与已知类别相似的目标预测更高的分数。
数据集:本文在MS-COCO(Microsoft Common Objects in Context)[28]数据集以及自行构建的化学实验器材检测数据集上评估模型的性能。MS-COCO是广泛应用的图像识别、分割和物体检测数据集,它包含了大量复杂且多样的日常生活场景图片,每张图片中标注了多个物体类别的信息。MS-COCO的数据划分方式与ORE[2]数据集的相同,将80个MS-COCO图像类别分成4个任务,每个任务有20个类实例,不与其他任务重叠,确保每个类别仅出现在一个任务中,将这4个任务记为。当学习时,将其中包括的所有类别视为已知,其他任务中的类别视为未知,学习其他任务时的情况以此类推。由于任务不存在未知类别,本文仅在前三个任务上进行方法性能评估。化学实验器材数据集包含了2 441张图像,涵盖了中学化学实验中需要用到的大部分实验仪器,包括pH试纸、比色卡、玻璃板、木塞、搅拌棒、试管、酒精灯和试剂瓶等(11种)类别,通过在这个数据集上的测试,评估模型的泛化能力,确保其在面对新类别时仍能保持较高的识别准确率。
YANGL P, XINT. Artificial intelligence promotes the transformation of educational assessment paradigm: Opportunities and challenges[J]. Journal of China Examinations, 2022(10): 13-21. DOI: 10.19360/j.cnki.11-3303/g4.2022.10.002(Ch ).
[3]
JOSEPHK J, KHANS, KHANF S, et al. Towards open world object detection[DB/OL]. [2024-01-23]. DOI: 10.1109/cvpr46437.2021.00577 .
[4]
GANINY, LEMPITSKYV. Unsupervised domain adaptation by backpropagation[J].Proceedings of Machine Learning Research, 2015,37: 1180-1189.
POGGIOT, GIROSIF. Networks for approximation and learning[J]. Proceedings of the IEEE, 1990, 78(9): 1481-1497. DOI: 10.1109/5.58326 .
[7]
RENS Q, HEK M, GIRSHICKR, et al. Faster R-CNN: Towards real-time object detection with region proposal networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6): 1137-1149. DOI: 10.1109/TPAMI.2016.2577031 .
[8]
ZHAOX W, MAY Q, WANGD R, et al. Revisiting open world object detection[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2024, 34(5): 3496-3509. DOI: 10.1109/TCSVT.2023.3326279 .
[9]
GUPTAA, NARAYANS, JOSEPHK J, et al. OW-DETR: Open-world detection transformer[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). New York: IEEE Press, 2022: 9225-9234. DOI: 10.1109/CVPR52688.2022.00902 .
[10]
CARIONN, MASSAF, SYNNAEVEG, et al. End-to-end object detection with transformers[C]//European Conference on Computer Vision. Cham: Springer, 2020: 213-229.10.1007/978-3-030-58452-8_13. DOI: 10.1007/978-3-030-58452-8_13 .
[11]
MAAZM, RASHEEDH, KHANS, et al. Class-agnostic object detection with multi-modal transformer[C]//European Conference on Computer Vision. Cham: Springer, 2022: 512-531.10.1007/978-3-031-20080-9_30. DOI: 10.1007/978-3-031-20080-9_30 .
[12]
LIUX F, YOOC, XINGF X, et al. Deep unsupervised domain adaptation: A review of recent advances and perspectives[J]. APSIPA Transactions on Signal and Information Processing, 2022, 11(1):e25. DOI: 10.1561/116.00000192 .
[13]
HOYERL, DAID X, VAN GOOLL. DAFormer: improving network architectures and training strategies for domain-adaptive semantic segmentation[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2022: 9914-9925. DOI: 10.1109/CVPR52688.2022.00969 .
[14]
LEED H. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks[DB/OL].[2024-02-21].
[15]
YANGY C, SOATTOS. FDA: Fourier domain adaptation for semantic segmentation[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). New York: IEEE Press, 2020: 408-4094. DOI: 10.1109/CVPR42600.2020.00414 .
[16]
ZHANGP, ZHANGB, ZHANGT, et al. Prototypical pseudo label denoising and target structure learning for domain adaptive semantic segmentation[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). New York: IEEE Press, 2021: 12414-12424. DOI: 10.1109/CVPR46437.2021.01223 .
[17]
MELAS-KYRIAZIL, MANRAIA K. PixMatch: Unsupervised domain adaptation via pixelwise consistency training[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). New York: IEEE Press, 2021: 12430-12440. DOI: 10.1109/CVPR46437.2021.01225 .
[18]
ZHOUQ Y, FENGZ Y, GUQ Q, et al. Context-aware mixup for domain adaptive semantic segmentation[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2023, 33(2): 804-817. DOI: 10.1109/TCSVT.2022.3206476 .
[19]
SOHNK, BERTHELOTD, LIC L, et al. FixMatch: Simplifying semi-supervised learning with consistency and confidence[EB/OL]. 2020: arXiv: 2001.07685.
[20]
GANINY, USTINOVAE, AJAKANH, et al. Domain-adversarial training of neural networks[EB/OL]. 2015: arXiv: 1505.07818.
SAITOK, WATANABEK, USHIKUY, et al. Maximum classifier discrepancy for unsupervised domain adaptation[C]//2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2018: 3723-3732. DOI: 10.1109/CVPR.2018.00392 .
[23]
ZHANGY M, GAOY J, LIH L, et al. Crucial semantic classifier-based adversarial learning for unsupervised domain adaptation[C]//2023 International Joint Conference on Neural Networks (IJCNN). New York: IEEE Press, 2023: 1-8. DOI: 10.1109/IJCNN54540.2023.10191498 .
ACUNAD, ZHANGG J, LAWM T, et al. F-domain-adversarial learning: Theory and algorithms[EB/OL]. 2021: arXiv: 2106.11344.
[26]
NGUYENX, WAINWRIGHTM J, JORDANM I. Estimating divergence functionals and the likelihood ratio by convex risk minimization[J]. IEEE Transactions on Information Theory, 2010, 56(11): 5847-5861. DOI: 10.1109/TIT.2010.2068870 .
[27]
JASTRZEBSKIS, SZYMCZAKM, FORTS, et al. The break-even point on optimization trajectories of deep neural networks[EB/OL]. 2020: arXiv: 2002.09572.
[28]
FORETP, KLEINERA, MOBAHIH, et al. Sharpness-aware minimization for efficiently improving generalization[EB/OL]. 2020: arXiv: 2010.01412.
[29]
LINT Y, MAIREM, BELONGIES, et al. Microsoft COCO: Common objects in context[M]//Computer Vision—ECCV 2014. Cham: Springer International Publishing, 2014: 740-755. DOI: 10.1007/978-3-319-10602-1_48 .
[30]
JINY, WANGX M, LONGM S, et al. Minimum class confusion for versatile domain adaptation[M]//Computer Vision—ECCV 2020. Cham: Springer International Publishing, 2020: 464-480. DOI: 10.1007/978-3-030-58589-1_28 .
[31]
LIUX Y, YUEZ Q, HUAX S. Proposal-level unsupervised domain adaptation for open world unbiased detector[EB/OL]. 2023: arXiv: 2311.02342.