Sliding window chaining of predicted camera poses

[13761152051]

Thank you very much for the author's work!!

I want to do an iterative approach to complete scene reconstruction, but I am currently facing issues with camera_poses.

My first inference used the following [00.jpg, 01.jpg, 02.jpg, 03. jpg], which would output [00_pose, 01_pose, 02_pose, 03_pose].

When I used Multi Modal Inference [02.jpg+02_pose, 03.jpg+03_pose, 04.jpg, 05. jpg] for my second inference , what I hope is that [02.jpg+02_pose] outputs 02_pose, so that it can be embedded into the result of the first inference

But it seems that the second output [02. jpg+02_pose] will be changed to an identity matrix, which means taking this point as the world coordinate origin

Is this the only output allowed by the model?

[Nik-V9]

Hi, yes, this is the only output allowed by the design of the model.

An easy way of getting all the subsequent poses in the coordinate frame of the first chunk is to apply the pose of 02_pose in the first inference to all the poses in the second chunk inference. This will achieve the sliding window alignment you are interested in.