Description
Thanks for the great work! While I have a bit confusion when reading your paper:
It is a bit unclear to me regarding the discussion and tables in supplmentary about Vloc tasks: accroding to the Tab.6, the performance degradation seems to come from the absence of GT intrinsics, wheras in the discussion it looks to point toward the sparsity of Cambridge SfM model?
Continue on the above question, have you tried the evalutaion with estimated instrinsics following "2D-matching" based loc method? A bit curious how that deveriate from the reported ones given GT intrinsics :D I would assume on 7scenes, roughly same performance based on "scaled rel-pose" method can be expected in theory?
From the code point of view
I would assume the implementation in dust3r/visloc.py is for "2D matching" based method. If so, can I understand below line of code as a "correction" of 2D projection based on GT intrinsics of query image?
Line 119 in a53d073
Do you mind also providing the implementation for "scaled rel pose" based loc solution?