Among solutions to the tasks of indoor localization and reconstruction, compared with traditional SLAM (Simultaneous Localization And Mapping), learning-based VO (Visual Odometry) has gained more and more popularity due to its robustness and low cost. However, the performance of existing indoor deep VOs is still limited in comparison with their outdoor counterparts mainly owing to large areas of textureless regions and complex indoor motions containing much more rotations. In this paper, the above two challenges are carefully tackled with the proposed SEOVO (Semantic Epipolar-constrained Online VO). On the one hand, as far as we know, SEOVO is the first semantic-aided VO under an online adaptive framework, which adaptively reconstructs low-texture planes without any supervision. On the other hand, we introduce the epipolar geometric constraint in an implicit way for improving the accuracy of pose estimation without destroying the global scale consistency. The efficiency and efficacy of SEOVO have been corroborated by extensive experiments conducted on both public datasets and our collected video sequences.