摘要
为解决图像到视频跨域物体检测任务中存在的视频运动模糊、目标遮挡姿态变化及图像与视频帧域偏移的问题,提出了一个新颖的图像到视频跨域物体检测框架。首先采用循环一致性对抗生成网络,将源域图像映射到目标视频域,生成与视频帧类似的图像(即伪视频帧),用于训练物体检测器。为进一步提升检测器的鲁棒性,提出时序传递模块,将检测结果传递到相邻视频帧以减少假阴性检测。同时设计上下文抑制模块,降低低置信度类别的分数,从而减少假阳性检测。在Youtube-Objects数据集上的实验结果验证了所提方法的有效性。
To solve the problems of motion blur,defocus and pose variations in videos and domain shift between image and video frames in image-to-video cross domain object detection,a novel image-to-video object detection framework was proposed.Firstly,a circular consistency countermeasure generation network was used to map the source domain image to the target video domain,and generate an image similar to the video frame(i.e.,a pseudo video frame)for training the object detector.To further improve the robustness of the detector,temporal propagation was employed to reduce false negatives by locally propagating detection results to adjacent frames and contextual suppression to reduce false positives by suppressing low-confidence classes.Experiments on the Youtube-Objects dataset demonstrate the effectiveness of the proposed method.
作者
闻子涵
陈谨
吴心筱
WEN Zihan;CHEN Jin;WU Xinxiao(School of Computer Science&Technology,Beijing Institute of Technology,Beijing 100081,China)
出处
《中国科技论文》
CAS
北大核心
2022年第11期1188-1193,1200,共7页
China Sciencepaper
基金
国家自然科学基金资助项目(62072041)。
关键词
深度学习
视频物体检测
域适应
时序上下文约束
deep learning
video object detection
domain adaptation
temporal and contextual constraint
