基于显式运动建模的视频伪装目标检测
Explicit motion handling for video camouflaged object detection
肖涛 1章超 2傅可人3
作者信息
- 1. 四川大学计算机学院,成都 610065
- 2. 四川警察学院,泸州 646000;智能警务四川省重点实验室,泸州 646000
- 3. 四川大学计算机学院,成都 610065;四川大学视觉合成图形图像技术国家级重点实验室,成都 610064
- 折叠
摘要
目前的视频伪装目标检测方法通常采用隐式运动建模或直接输入存在噪声的离线光流图来获取运动线索,这会影响模型性能.为了解决这一问题,提出一种新的基于显式运动建模的视频伪装目标检测框架,称为SMHNet.首先,该框架将显式运动建模与伪装目标检测联合在同一个框架中进行学习.然后利用特征双向更新模块实现两个分支的双向交互更新,相互补充、优化和纠错,输出光流估计结果和目标检测图.此外,为了解决缺少光流真值图这一问题,采用自监督策略对显式运动建模分支进行监督.在两个数据集上的对比实验结果表明,SMHNet有效地提高了视频场景中伪装目标检测的性能.
Abstract
Earlier video camouflaged object detection methods often exploit motion cues by implicit motion modeling or directly inputting offline optical flow maps with noise,which affects model performance.To address the effective utilization of motion cues,an explicit motion modeling framework for video camouflaged object detection,called SMHNet,was proposed.First,an explicit motion modeling branch and a camouflaged object detection branch were jointly learned in the same framework.Then,the two branches were updated bidirectionally using a bidirectional feature updating module.The two branches performed mutual optimization and error correction to output optical flow estimation results and object detection maps.In addition,to address the lack of ground truth optical flow,a self-supervised strategy was adopted to supervise the explicit motion modeling branch.Comparison experiments on two benchmark datasets show that SMHNet effectively improves the performance of video camouflaged object detection.
关键词
视频伪装目标检测/显式运动建模/光流/自监督Key words
Video camouflaged object detection/explicit motion handling/optical flow/self-supervision引用本文复制引用
基金项目
国家自然科学基金(62176169)
智能警务四川省重点实验室开放基金(ZNJW2022KFMS001)
出版年
2024
NETL
NSTL
维普
万方数据