Hi, @lyclyc52 @zymk9, could you help me? As part of my experiment, I would like to test different NeRF models that produce more than just sigma and RGB for object detection on Front3D (let's start from this). I don't know how to sample new features without knowing the scene H, W, L and resolution step parameter in the way to be consistent with 3d bbox annotations.
1) I find in front3d_nerf_data.scene_i.transforms.json field room_box. Is it min max from which i can calculate H, W, L similar to this?
2) How to deal with the resolution step? I thought it is one, but after this i am not sure. I suspect scale field in transforms.json is somehow related too ...
Since we do not have the ground truth scene bounding boxes of scannet, we use that function to estimate the scene bounding box based on the object bounding box. For Front3D, we have the ground truth scene boxes so we do not have to do so.
This is to resize the whole scene to the targeted resolution. We resize it based on the longest length of the scene bounding box. In other word, after resizing, the longest length = args.max_res.
Hi, @lyclyc52 @zymk9, could you help me? As part of my experiment, I would like to test different NeRF models that produce more than just sigma and RGB for object detection on Front3D (let's start from this). I don't know how to sample new features without knowing the scene H, W, L and resolution step parameter in the way to be consistent with 3d bbox annotations.
1) I find in front3d_nerf_data.scene_i.transforms.json field room_box. Is it min max from which i can calculate H, W, L similar to this? 2) How to deal with the resolution step? I thought it is one, but after this i am not sure. I suspect scale field in transforms.json is somehow related too ...
Appreciate any help, thx.