What improvements have been made to the orbslam system

[molong-d]

What improvements have been made to the orbslam system Author, thank you for sharing your code. I have added your code to the orbslam3 system and successfully implemented it, but the overall trajectory accuracy has not changed much. On some sequences in the TUM dataset, there is a slight increase in accuracy, while others have a slight decrease. May I ask why this is? Or can this boot enhance which part of the orbslam3 system? I hope for a reply, thank you very much. @Antu3heng @SJTU-ViSYS

[Antu3heng]

We didn't test the modified ORB SLAM2 in the TUM dataset. The results in the EuRoc dataset can be found in the appendix of our paper. For the TUM dataset, the images suffered from the motion blur issue which is not the main data of our training data, maybe it can explain the performance in the TUM dataset.

As for the improvement, I think the boosted ORB can help the ORB-SLAM find more correspondence to make the pose estimation more robust. However, the frontend of ORB SLAM is inherently robust, as it can find enough correspondence through methods like reprojection. Therefore, even the improvement brought by end-to-end deep learning methods such as SuperPoint to ORB SLAM is very limited.

In fact, I think the greatest benefit of learning-based local features so far is the ability to create a more complete visual feature map that is less affected by changes in viewpoint and lighting conditions, which can be utilized for loop closure and subsequent relocalization tasks.

[Cmomo1113]

您好，可以请问以下把这个代码添加在orb-slam3的系统中应该改哪些地方吗 ？希望得到回复，非常感谢！@molong-d @Antu3heng @SJTU-ViSYS

[guker]

We didn't test the modified ORB SLAM2 in the TUM dataset. The results in the EuRoc dataset can be found in the appendix of our paper. For the TUM dataset, the images suffered from the motion blur issue which is not the main data of our training data, maybe it can explain the performance in the TUM dataset.

As for the improvement, I think the boosted ORB can help the ORB-SLAM find more correspondence to make the pose estimation more robust. However, the frontend of ORB SLAM is inherently robust, as it can find enough correspondence through methods like reprojection. Therefore, even the improvement brought by end-to-end deep learning methods such as SuperPoint to ORB SLAM is very limited.

In fact, I think the greatest benefit of learning-based local features so far is the ability to create a more complete visual feature map that is less affected by changes in viewpoint and lighting conditions, which can be utilized for loop closure and subsequent relocalization tasks.

I also test superpoint on slam mapping, the result shows no gain.