I have two questions regarding reproducing pairs from the datasets for training.
Habitat
Are the scenes the pairs are generated from the same as in CroCo Habitat README?
Specifically, do the 1M pairs relate in some way to the ~1.8M pairs used in Croco?
Real datasets
For CroCoV2 you provide metadata to re-generate the crops for ARKitScenes and MegaDepth.
Specifically the CroCoV2 paper mentions,
1,070,414 pairs from ARKitScenes [8], 2,014,789 pairs from MegaDepth
Do these relate to the pairs you obtained for dust3r training in some way?
If there is some metadata similar to the one for CroCo to generate these pairs that would be greatly appreciated!
First of all: very cool work!
I have two questions regarding reproducing pairs from the datasets for training.
Habitat
Are the scenes the pairs are generated from the same as in CroCo Habitat README? Specifically, do the 1M pairs relate in some way to the ~1.8M pairs used in Croco?
Real datasets
For CroCoV2 you provide metadata to re-generate the crops for ARKitScenes and MegaDepth. Specifically the CroCoV2 paper mentions,
Do these relate to the pairs you obtained for dust3r training in some way?
If there is some metadata similar to the one for CroCo to generate these pairs that would be greatly appreciated!
Thank you!