Charles R. Qi

Hi @KleinYuan Great effort in building the server! The py_func for `compute_box3d_iou` shouldn't matter for inference because it is only used for computing the loss. There has been some third-party...

One way to work around is to directly element-wise multiply the mask with the points. It increases the memory usage a bit, but won't affect performance. It should also be...

HI @lyp0413 Those are all hyperparameters for the model. We didn't tune those numbers too much but did try a few options (like 3 options) for each of them. I...

@jytime I think your training time is reasonable. If you are using SSD as well, then I believe the GPU costs the difference.

Hi @charlesCXK The major change is the learning rate scheduling. Originally we have exponential decay of the learning rate for paper version-1. Later we updated it to step-wise learning rate...

@rttariverdi67 Have you used the V1.3 Waymo Open Dataset or the old version? The 3d semseg labels only exist in V1.3. Regarding the annotated frames, we only have 3d semseg...

@SizheWei I believe we have 3d semseg labels on every run segment/tfrecord file. To find out which frames have labels, yes, you'll need to iterate over the frames in each...