EFRNet-VL: An end-to-end feature refinement network for monocular visual localization in dynamic environments

Jingwen Wang, Hongshan Yu, Xuefei Lin, Zechuan Li, Wei Sun, Naveed Akhtar

Abstract: Highlights•Proposed an end-to-end deep learning model to estimate camera pose in real time.•EFRNet-VL can induce feature invariance to drastic condition variations.•It uses self-attention and LSTMs modules to refine learning-based features.•Validated on three well-known public datasets, including indoor and outdoor scenes.•Demonstrated superiority of EFRNet-VL in various challenging conditions.