作者:李成严，郑企森，王昊
Authors:LI Chengyan,ZHENG Qisen,WANG Hao
摘要:针对像素级自适应较大的图像翻译偏差,特征级自适应的源偏判别风险以及弱监督学习无法兼顾检测准确性和实时性等问题,提出了多元化域移位器和伪边界框生成器以逐步调整预训练模型,在像素级与特征级渐进完成自适应的域迁移框架 。通过域移位器从源域生成多样化的中间域图像调整检测模型以弥合域差距 ,减小图像翻译偏差 。将中间域作为监督的源域,并结合目标域中的图像级标签生成伪标注图像调整检测模型以改善源偏判别性 。基于 SSD算法构建与域迁移框架相匹配的实时目标检测器,实现弱监督条件下的实时目标检测 。在 PASCAL VOC 迁移至 Clipart1k 等数据集上的 mAP 优于现有方法 0. 4% ~ 4. 7% ,检测速度为 32 FPS ~ 47 FPS,提高准确率的同时满足了实时检测的要求 ,具有更优越的迁移检测性能。
Abstract:Aiming at the problems of large image translation bias at the pixel-level adaptation, the risk of source-bias discrimination at the feature-level adaptation, and the inability of weakly supervised learning to balance detection accuracy and real- time performance, a diversified domain shifter and pseudo bounding box generator are proposed to gradually adjust the pre-training model. The adaptive cross-domain framework is gradually completed at pixel-level and feature-level. A diversified intermediate domain adjustment detection model is generated from the source domain by a domain shifter to bridge the domain gap and reduce the image translation bias. The intermediate domain is used as the supervised source domain, and the pseudo-labeled image adjustment detection model is generated by combining image-level annotations in the target domain to improve source-bias discrimination. A real-time object detector matching the cross-domain framework is constructed based on SSD algorithm to realize real-time object detection under weakly supervised conditions. The mAP on PASCAL VOC migrated to Clipart1k and other datasets is 0. 4% ~ 4. 7% better than the existing methods. The detection speed is 32 FPS ~47 FPS. This improves the accuracy and meets the requirements of real-time detection, and has better migration detection performance.

PDF全文下载地址:

可免费Download/下载PDF全文