Insights on Evaluation of Camera Re-localization Using Relative Pose Regression

We consider the problem of relative pose regression in visual relocalization.\nRecently, several promising approaches have emerged in this area. We claim that\neven though they demonstrate on the same datasets using the same split to train\nand test, a faithful comparison between them was not available since on\ncurrently used evaluation metric, some approaches might perform favorably,\nwhile in reality performing worse. We reveal a tradeoff between accuracy and\nthe 3D volume of the regressed subspace. We believe that unlike other\nrelocalization approaches, in the case of relative pose regression, the\nregressed subspace 3D volume is less dependent on the scene and more affect by\nthe method used to score the overlap, which determined how closely sampled\nviewpoints are. We propose three new metrics to remedy the issue mentioned\nabove. The proposed metrics incorporate statistics about the regression\nsubspace volume. We also propose a new pose regression network that serves as a\nnew baseline for this task. We compare the performance of our trained model on\nMicrosoft 7-Scenes and Cambridge Landmarks datasets both with the standard\nmetrics and the newly proposed metrics and adjust the overlap score to reveal\nthe tradeoff between the subspace and performance. The results show that the\nproposed metrics are more robust to different overlap threshold than the\nconventional approaches. Finally, we show that our network generalizes well,\nspecifically, training on a single scene leads to little loss of performance on\nthe other scenes.\n