Scene-agnostic Pose Regression for Visual Localization

Figure 1. 360SPR panoramic dataset for Scene-agnostic Pose Regression (SPR). Trajectories with different lengths and sampling intervals are collected at 3 heights of 360°sensors, i.e., 0.1m, 0.5m, 1.7m, corresponding to sweeping , quadruped , and humanoid robots. The proposed SPR model trained from known scenes can generalize well to unknown scenes, without database retrieval.