Hi, I'm reading your NeurIPS 2021 paper "Voxel-based 3D Detection and Reconstruction of Multiple Objects from a Single Image". I have some doubts as following:
(1) How to output 3d heatmap and 3d regeression using 2d convs on a 4d tensor?
(2) What's the function of 3d u-net? Is there any ablation study on this?
Looking forward to your reply. Thank you.
Hi,
(1) We lift the 2D feature to 3D space, (code). For the implementation of the 3D heatmap and 3D regression, please find here.
(2) Basically, we believe 3D U-net could integrate both local and global information, but we didn't quantitatively ablate that.
Hi, I'm reading your NeurIPS 2021 paper "Voxel-based 3D Detection and Reconstruction of Multiple Objects from a Single Image". I have some doubts as following: (1) How to output 3d heatmap and 3d regeression using 2d convs on a 4d tensor? (2) What's the function of 3d u-net? Is there any ablation study on this? Looking forward to your reply. Thank you.