研究简介: 深度双目模型在驾驶场景中取得了最先进的性能,但在未见过的场景中进行测试时性能严重下降。尽管最近的工作通过不断的在线适应缩小了这一性能差距,但这种设置需要在部署时不断更新梯度,并且无法避免灾难性的遗忘。为了应对这些挑战,我们建议执行连续双目匹配,其中模型的任务是 1) 不断学习新场景,2) 克服忘记先前学习的场景,以及 3) 在没有在线梯度更新的情况下连续预测视差。我们通过引入可重用架构增长 (RAG) 框架来实现这一目标。 RAG 利用特定任务的神经单元搜索和架构增长来持续学习新场景。在增长过程中,它可以通过重用之前的神经单元来保持高可重用性,同时获得良好的性能。还引入了一个名为 Scene Router 的模块,以在推理时自适应地选择特定于场景的架构路径。实验结果表明,我们的方法在各种具有挑战性的驾驶场景中都优于最先进的方法。 RAG模型结构示意
What is new
- Sep. 2023: 3 papers accepted NeurIPS'2023
- Mar. 2023: 9 papers accepted in CVPR'2023
- Mar. 2022: 11 papers accepted in CVPR'2022
- Mar. 2021: 7 papers accepted in CVPR'2021
- Feb. 2021: I will serve as the Area Chair of ACM MM'2021
- Feb. 2021: I will serve as the Associate Editor of the Journal of IJAC
- Nov. 2020: I will serve as the Area Chair of ICCV'2021
- Nov. 2020: I will serve as the Associate Editor of the journal of CJIG [中国图象图形学报]
- Jul. 2020: I will serve as the Area Chair of IJCAI'2021
- Jul. 2020: 3 papers accepted in ECCV'2020
- Jul. 2020: Glad to have my position promoted in the CEBSIT, CAS
- Apr. 2020: I am leading a “2035 Innovation Team”of AI fundermental research in CASIA
- Mar. 2020: 2 of the 5 papers selected as oral (top 5%) in CVPR'2020
- Feb. 2020: 5 papers accepted in CVPR'2020
- Jan. 2020: I will serve as the Area Chair of ACM MM'2020
- Jan. 2020: I will serve as the Area Chair of CVPR'2021
- Dec. 2019: I have been an Associate Editor of IEEE T-CSVT
- Sep. 2019: I have been an Associate Editor of Pattern Recognition