摘要
针对目前增量学习中所面向目标都是固定姿态这一现象,本文考虑了更严格的设定,即面向多姿态目标的在线类增量学习,并提出了无视姿态重放方法来缓解在线类增量学习中面对多姿态目标时的灾难性遗忘。首先,将2D/3D目标进行点云化处理,以方便提取目标的有效几何信息;其次,基于SE(d)(d=2,3)群对网络进行平移旋转等变性改进,使网络能够提取更丰富的几何信息,从而降低模型在每个任务中受目标姿态的影响;最后,根据损失变化采样特定样本用于重放来缓解灾难性遗忘。实验结果表明,在面对固定姿态目标MNIST、CIFAR-10时,本文方法的最终平均精度分别达到了88%和42.6%,与对比方法结果相近,但最终平均遗忘率明显优于对比方法,分别降低了约3%和15%。在面对多姿态目标RotMNIST、trCIFAR-10时,本文方法依旧能很好地保持在固定姿态目标中的表现,基本不受目标姿态的影响。此外,在3D数据集ModelNet40中的表现也依旧稳定。本文所提方法在在线类增量学习中能够不受目标姿态的影响,同时能缓解灾难性遗忘,具有很好的稳定性和可塑性。
In response to current phenomenon that all targets in incremental learning are fixed pose,this paper considers a more rigorous setting,i.e.online class incremental learning for multi-pose targets,which innovatively proposes an ignoring pose replay method to alleviate the catastrophic forgetting in facing multi-pose targets in online class incremental learning.Firstly,2D/3D targets are point-clouded to facilitate the extraction of useful geometric information.Secondly,the network modifies for equivariance based on the SE(d)(d=2,3)group to enable the network to extract richer geometric information,thus reducing the impact of target poses on the model in each task.Finally,specific samples are sampled for replay to mitigate catastrophic forgetting based on loss variation.Experimental results show that when facing fixed posture targets MNIST and CIFAR-10,final average accuracy reaches to 88%and 42.6%respectively,which is comparable to the comparison method,and final average forgetting is significantly better than the comparison method,with a reduction of about 3%and 15%respectively.In the case of the multi-pose target RotMNIST and trCIFAR-10,the proposed method continues to perform well in fixed-pose targets,largely independent of target pose.In addition,the performance in 3D datasets ModelNet40 and trModelNet40 remains stable.The method proposed is able to be independent of the target pose in online class incremental learning,while achieving catastrophic forgetting mitigation,with excellent stability and plasticity.
作者
张润江
郭杰龙
俞辉
兰海
王希豪
魏宪
ZHANG Run-jiang;GUO Jie-long;YU Hui;LAN Hai;WAGN Xi-hao;WEI Xian(College of Electrical Engineering and Automation,Fuzhou University,Fuzhou 350108,China;Fujian Institute of Research on the Structure of Matter,Chinese Academy of Sciences,Fuzhou 350108,China;Quanzhou Institute of Equipment Manufacturing,Haixi Institutes,Chinese Academy of Sciences,Quanzhou 362000,China)
出处
《液晶与显示》
CAS
CSCD
北大核心
2023年第11期1542-1553,共12页
Chinese Journal of Liquid Crystals and Displays
基金
福建省科技计划(No.2021T3003)
泉州市科技计划(No.2021C065L)。
关键词
在线类增量学习
灾难性遗忘
无视姿态重放
等变性
点云分类
online class-incremental learning
catastrophic forgetting
ignoring pose replay
equivariance
point cloud classification
