You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Your results showed that you use NeRF dataset as input in ImVoxelNet and FCAF3D, but these two models use point cloud as input. But the feature of point cloud and NeRF maybe different, which could cause the original backbone to be useless. Did you modified code of these two models? If so, could you please offer the code you have modified?
The text was updated successfully, but these errors were encountered:
We do not modified the code of these two models and we only slightly change the input. As described in the paper, "FCAF3D is a point cloud-based 3D detector, so we use the ground-truth depth from Hypersim and ScanNet, NeRF rendered depth for 3D-FRONT, and the corresponding RGB images to build point clouds. "
Your results showed that you use NeRF dataset as input in ImVoxelNet and FCAF3D, but these two models use point cloud as input. But the feature of point cloud and NeRF maybe different, which could cause the original backbone to be useless. Did you modified code of these two models? If so, could you please offer the code you have modified?
The text was updated successfully, but these errors were encountered: