Thanks for your public work.
I want to ask how robust the model if I use it this way:
- Carefully precapture N frames of a bounded scene, cover everything in that scene.
- Frames are captures in a way that there is good amount of overlap between t-1 and t.
- Use Direct Mode, set keyframe_interval = 1 so all these N frames get its full KV cache stored.
- New frame arrives (Let call it Query frame), taken from random position of the scene, not overlap with the last frame t_N of N frames above, but definitely overlap with at least one of them.
- Get the Camera Pose of Query frame.
In short, I want to use lingbot-map purely as Visual Localization. Will using it that way degrade performance too much?
As I understand it will lose the Local Pose-reference Window advantage, where the overlap of Query frame is with somewhere in the N frames of KV cache, but in this case not with the latest frame t_N, so the time order is messed up.
I tested that way of use, and the camera pose seems quite good to my human eyes, but I dont have a way to measure it exactly in metric.
Thanks for reading.
Thanks for your public work.
I want to ask how robust the model if I use it this way:
In short, I want to use lingbot-map purely as Visual Localization. Will using it that way degrade performance too much?
As I understand it will lose the Local Pose-reference Window advantage, where the overlap of Query frame is with somewhere in the N frames of KV cache, but in this case not with the latest frame t_N, so the time order is messed up.
I tested that way of use, and the camera pose seems quite good to my human eyes, but I dont have a way to measure it exactly in metric.
Thanks for reading.