Isn't Visual Odometry just Ego-Motion? Why separate them?

uzh-rpg / event-based_vision_resources

2.87k stars 651 forks source link

Isn't Visual Odometry just Ego-Motion? Why separate them? #256

Closed DongHappyyy closed 9 months ago

DongHappyyy commented 9 months ago

My understanding is that the Visual Odometry is a Ego-Motion (also called Camera Motion). Is there a misunderstanding in my understanding?

guillermogb commented 9 months ago

The SLAM (Simultaneous Localization And Mapping) section has several subsections, such as:

Localization, Ego-motion estimation -> estimate camera motion given a map of the scene.
Visual Odometry -> no map is given. Both, scene map and motion are estimated concurrently.
etc. This is what, I believe, is a common meaning for these subsection titles.

DongHappyyy commented 9 months ago

Can I understand that within the scope of SLAM, there is a difference between visual odometry and ego motion (map is given or not); If SLAM is not involved, then VO and ego motion are equivalent.

guillermogb commented 9 months ago

Maybe I am a bit relaxed about the exact meaning of these keywords. You would have to look them up more precisely in a glossary about computer vision or robotics. What's the concrete suggestion or question about the list in the repository?

DongHappyyy commented 9 months ago

I saw a sentence in the paper (Part I: The First 30 Years and Fundamentals) written by Professor Davide Scaramuzza: "Visual odometry (VO) is the process of estimating the egomotion of an agent (e.g., vehicle, human, and robot) using only the input of a single or multiple cameras attached to it.", so my understanding is that Visual Odometry and Ego-motion estimation are essentially the same thing, just with different names.

In another article titled "Deep learning based robust positioning for all weather autonomous driving", it was mentioned that "Ego motion estimation (also called odometry) with onboard sensors provides a complete localization solution in challenging environments, predicting the accurate relative self position of AVs." This further confirms my understanding.

I have no specific suggestions, on the contrary, I respect your summary deeply. I just see that you separated VO and ego motion, and I think my understanding is incorrect, so I want to communicate with you.

guillermogb commented 9 months ago

Would you be satisfied if "Localization, Ego-motion estimation" was replaced by "Localization, pose estimation"? https://github.com/uzh-rpg/event-based_vision_resources/pull/258 ?
What "summary" do you refer to? (the repository?)

DongHappyyy commented 9 months ago

It's good to keep it for now. To be honest, I didn't think of a particularly good classification method either. Pose estimation is not very appropriate either. As far as I know, there is still a gap between pose estimation (more used in SfM ) and motion estimation (more used in SLAM)

guillermogb commented 9 months ago

Alright. Then we can close this issue and delete the PR. These kind of discussions are usually communicated via e-mail, I think.