OFVL-MS: Once for Visual Localization across Multiple Indoor Scenes

• DOI: 10.48550/arxiv.2308.11928
• Publication Date: 2023-01-01
• Authors: Tao Xie; Kun Dai; Siyi Lu; Ke Wang; Zhiqiang Jiang; Jinghan Gao; Dedong Liu; Jie Xu; Lijun Zhao; Ruifeng Li
• Keywords: Pace; Normalization

In this work, we seek to predict camera poses across scenes with a multi-task learning manner, where view the localization of each scene as new task. We propose OFVL-MS, unified framework that dispenses traditional practice training model for individual and relieves gradient conflict induced by optimizing multiple collectively, enabling efficient storage yet precise visual all scenes. Technically, in forward pass design layer-adaptive sharing policy learnable score layer automatically determine whether is shared or not. Such empowers us acquire task-shared parameters reduction cost task-specific scene-related features alleviate conflict. backward introduce normalization algorithm homogenizes magnitude so tasks converge at same pace. Furthermore, sparse penalty loss applied on scores facilitate parameter without performance degradation. conduct comprehensive experiments benchmarks our released indoor dataset LIVL, showing OFVL-MS families significantly outperform state-of-the-arts fewer parameters. also verify can generalize much few while gaining superior performance.