The domain gap between dark scenes and the data used by traditional pretrained models leads to suboptimal performance with the conventional pretrain-finetune approach,and pretraining from scratch is costly.To address this issue,a domain-adaptive pretraining method is proposed to improve action recognition performance in the dark environments.The method integrates an external vision enhancement model for de-darkening to introduce critical knowledge for dark scene processing.It also employs a cross-domain self-distillation framework to reduce the domain gap of visual representations between illuminated and dark scenes.Through extensive experiments in various dark environment action recognition settings,the proposed approach can achieve a Top1 accuracy of 97.19%on the dark dataset of fully supervised action recognition.In the source-free domain adaptation on the Daily-DA dataset,the accuracy can be improved to 49.11%.In the multi-source domain adaptation scenario on the Daily-DA dataset,the Top1 accuracy can reach 54.63%.
dark scenesaction recognitiontransfer learningdomain adaptation
维普
万方数据
