Hello, Thanks for this great job. I would like to make a comment on PointPillars repo forked from this source repo. In pointpillars I see it is used point pillar feature extraction layer instead of Voxelnet and it directly goes to the 2D convolution as in paper. But this 2d detection layer is still RPN like in voxelnet not SSD like in original paper. It may be an important point to obtain real pointpillars architecture results and wanted to ask if Im wrong.
Hello, Thanks for this great job. I would like to make a comment on PointPillars repo forked from this source repo. In pointpillars I see it is used point pillar feature extraction layer instead of Voxelnet and it directly goes to the 2D convolution as in paper. But this 2d detection layer is still RPN like in voxelnet not SSD like in original paper. It may be an important point to obtain real pointpillars architecture results and wanted to ask if Im wrong.