A question regarding demo_inference.py

[ZhHaoH]

Dear Author, thank you for your great work. I have a minor question regarding the relative pose estimation results obtained using the Mickey model provided in your demo_inference.py. Specifically, when estimating two frames from the validation set of the map-free dataset, the relative pose calculated through submission.py shows good visual alignment with ground truth after visualization. However, the results directly obtained from demo_inference.py demonstrate significant visual discrepancies from the ground truth. Could you please explain the potential reasons for this inconsistency?

[axelBarroso]

Hello, both scripts are using the same pose estimation code (see here and here). The method has some probabilistic steps, but I would not expect the results to be very different. Thus, results should be comparable.

Some things that one could check to validate the results:

1. I would check i the intrinsics are given correctly to the demo script.
2. Order of the images matter, are images given in the desired order?
3. Are you using the same checkpoints in both scripts?

Thanks!