文献类型：期刊文献

英文题名：Opengcd: Assisting Open World Recognition with Generalized Category Discovery

作者：Gao, Fulin[1]; Zhong, Weimin[1]; Cao, Zhixing[1]; Peng, Xin[1]; Li, Zhi[1]

机构：[1] Key Laboratory of Smart Manufacturing in Energy Chemical Process, Ministry of Education, East China University of Science and Technology, Shanghai, 200237, China

年份：2024

外文期刊名：SSRN

收录：EI(收录号：20240307986)

语种：英文

摘要：A desirable open world recognition (OWR) system requires cycling through three subtasks: (1) Open set recognition, i.e., classifying the known (classes seen during training) and rejecting the unknown (unseen/novel classes) online; (2) Grouping and labeling these unknown as novel known classes; (3) Incremental learning (IL), i.e., continual learning these novel classes and retaining the memory of old classes. As a challenging task, existing OWR methods mostly assume that the second subtask is performed entirely by expensive human labor and employ exemplar replay to solve the third subtask. We observe that the existence of exemplars enables the implementation of generalized category discovery (GCD) in OWR. GCD can reduce the human cost of the second subtask by clustering unknown instances using exemplars as supervision. Thus, we propose a new benchmark of assisting open world recognition with generalized category discovery (OpenGCD). Compared to the prevalent OWR benchmark, OpenGCD turns the second subtask into GCD and minor manual labeling revision. Moreover, OpenGCD devises mutual information-based soft label assignment and memory allocation to address the problems that popular uncertainty calibration and exemplar replay methods suffer from raw image corruption and inefficient resource utilization when solving the first and third subtasks, respectively. Furthermore, we propose a new metric for GCD, called harmonic clustering accuracy, to provide a more sensible evaluation. Experimental and visualization results show that OpenGCD is feasible and state-of-the-art. Code: https://github.com/Fulin-Gao/OpenGCD. ? 2024, The Authors. All rights reserved.