adapt to nuscene dataset
Closed this issue · 6 comments
Hi, I'm trying to adapt Mars to nuscene dataset, I used the frontleft， front， frontright，three cameras to reconstruct the scene，but the result of the frontleft is a bit strange (like the following image). Do you have any suggestion about this? Thank you in advance！
Could you provide some more information?
can you open this link？wandb In this version，I used the v1.0-mini version dataset，I used frontleft and front two cameras，and the scene id is 0757，I used the first 15 frames to train Mars
just from the rendered object_rgb images, it can be seen that the object bounding boxes are too large so that some of the background is included in the bbox.
I reduced the size of the bounding box，but the rendering results are still strange， here is the link wandb，I'm confused whether it's because the camera and object pose of this dataset is too noisy？ or because the overlaps between adjacent camera views are too small to be effective for training NeRFs？
Yes, this is actually what we also concluded for the nuscenes dataset. The provided boxes & poses are a bit too noisy.
@wutongtong hi, I want to ask a question for you. Did you reconstruct the nuscene data succesfully? And how to code the dataparser for 6v cameras?