Open Sanyam-Mehta opened 1 year ago
In relocalization people often do image based search first to get images that are close to the query image. So in droid slam you need to do
loop closure is done by performing local ba if it finds previous frames inside the radius
@Sanyam-Mehta you need image based localization...or appearance based localization..check RTabMap
Are you referring to the timestamp radius when you mention "radius"? In that case, the inside radius would be related to odometry rather than loop closure, correct?
@kwea123
Let's say I have built a map using a dataset and saved the results on the disk. During the next step, I want to utilize the saved results to localize the position of the camera using a new frame. What could I do using the Droid Slam approach?
My understanding of the code tells me that during the map-building process, the pose of the next frame is kept equal to the estimated pose of the current frame and then it's optimized as follows:
Now the issue with reuse is that the first step that extracts the frames in the neighbourhood of the current frame uses the pose estimate of the last frame. For pure relocalization using a saved map, it is impossible to get this initial pose estimate. There is no feature-based matching step either that we could use to extract the frames close to the current frame. In such a scenario, what should be done to localize a frame using the saved map?
This is also related to the question of loop closure: How is it performed in the current setting?