Open-world Weakly-Supervised Object Localization

Jinheng Xie, Zhaochuan Luo, Rouyi Li, Yawen Huang, Haozhe Liu, Yuexiang Li, Yefeng Zheng, Yang Zhang, Linlin Shen, Mike Zheng Shou

Abstract: Highlights•Introduce the Open-world Weakly Supervised Object Localization (OWSOL) task.•Propose contrastive co-learning and generalized class activation mapping for OWSOL.•Re-organize three datasets to serve as evaluation benchmarks for OWSOL.•Experiments show that the proposed paradigm surpasses the baseline by a large margin.