Jianyuan Wang

In the 2-gpu setting, can you still try to import from /opt/conda/envs/vggt/lib/python3.10/site-packages/pycolmap.cpython-310-x86_64-linux-gnu.so ?

Hi if I remember correclty, you just need to set num_workers=0, and then you can set breakpoints in vscode, or use pdb for debugging

Hi, if I remember correctly, set_epoch() was designed for other experimental functions. It should not affect the training. Please let me know if you think there is a bug.

Hi currently I think this is the known limitation of the feedforward models, i.e., the alignment may have small mismatches. Did you try the reconstruction with the BA option enabled?...

Hi I cannot see the input images, while a wild guess is, probably this results from a kind of doppelgangers? https://doppelgangers25.github.io/doppelgangers_plusplus/

Hi @guangkaixu @jianglh-WHU , We didn’t follow the DUST3R strategy. Instead, we compute camera-pose-based similarity between frames using the code below. For each frame, we rank all others by pose...

Hi @Livioni , Camera tokens and register tokens are learnable tokens, they are defined inside the network. They will aggregate the information according to the input frames, and do not...

Hi @Livioni could you elaborate more about this? Sorry I am still confused about the question here. Hi @lifuguan , no, only those images from one scene (sequence/image collection/...) will...

It should be BxS ~= 48, and S is a random number between 2-24. For example, if S=12, B=4. If S=7, B=6.

HI @davnords , they are sampled from different scenes, while not necessary to be from the same dataset. I am cleaning the training code now, hope to share more this...