Hello,
The performance reported for Occ3D-Waymo raises a few questions, I'd be happy to have your return on this:
- Why do you not also report your CTF-Occ model from the Occ3d paper, which was SOTA ?
- Why is the performance wildly different from the Occ3d paper's results for the same task and dataset ? The reported values for BEVFormer in the Occ3d paper is 18.73mIoU, while in your CVT-Occ paper, it's 24.58 for the same model and task ? Even BEVFormer w/o TSA obtains a 23.87 mIoU. Even accounting for the "Motorcycle" class that was removed in this paper, the difference is still of more than 3 mIoU
- Why don't you evaluate on the 0.1m voxel size from Occ3D-Waymo as well ? It would be interesting to see.
Thank you in advance
Hello,
The performance reported for Occ3D-Waymo raises a few questions, I'd be happy to have your return on this:
Thank you in advance