Official PyTorch implementation of DD3D: Is Pseudo-Lidar needed for Monocular 3D Object detection? (ICCV 2021), Dennis Park*, Rares Ambrus*, Vitor Guizilini, Jie Li, and Adrien Gaidon.
Big progress, good architecture. The inference latency should also be reduced.
But I have two questions.
What's the density of 3d points outputted by the feature map? [Should as dense as image resolution, It doesn't matter.]
As we know, image is more dense compared to point cloud produced by lidar, even dense lidar. [Alright, we can sub-sample it latter.]
OK, state of the art 3d detector head can be used, such as center-point, it is efficient.
Hum, I think the only latency bottleneck may be the dense 3d point sub-sample part.
3d points sub-sample methods designed for Lidar generated data is in-efficient here, should be the bottleneck.
The output of the confidence could be used here. But, is the predicted confidence confidence?
Hope to see someone help to design good sub-sample method here.
Right, I have not read the code. Just a glimps of the paper. Lots of self guess.
Big progress, good architecture. The inference latency should also be reduced.
But I have two questions. What's the density of 3d points outputted by the feature map? [Should as dense as image resolution, It doesn't matter.] As we know, image is more dense compared to point cloud produced by lidar, even dense lidar. [Alright, we can sub-sample it latter.] OK, state of the art 3d detector head can be used, such as center-point, it is efficient.
Hum, I think the only latency bottleneck may be the dense 3d point sub-sample part. 3d points sub-sample methods designed for Lidar generated data is in-efficient here, should be the bottleneck. The output of the confidence could be used here. But, is the
predicted confidence
confidence? Hope to see someone help to design good sub-sample method here.Right, I have not read the code. Just a glimps of the paper. Lots of self guess.