search

tusen-ai / SST

Code for a series of work in LiDAR perception, including SST (CVPR 22), FSD (NeurIPS 22), FSD++ (TPAMI 23), FSDv2, and CTRL (ICCV 23, oral).
Apache License 2.0
801 stars 102 forks source link

When I change the seg_voxel_size to train custom dataset, the voxel2point neck gives the following error,what does it mean? #70

Closed XLMore closed 2 years ago

XLMore commented 2 years ago

image

Abyssaledge commented 2 years ago

It means you may not correctly mask those points out of point_cloud_range.

XLMore commented 2 years ago

It means you may not correctly mask those points out of point_cloud_range. I set the voxel_size and pc_range as follows: seg_voxel_size = (0.16, 0.16, 0.22) point_cloud_range = [-80, -32, -2.2, 80, 32, 4.4] How can i fix it?Thank you for your quick reply！

Abyssaledge commented 2 years ago

How about the point_cloud_range in your data_cfg? The one in data_cfg is used in PointsRangeFilter.

XLMore commented 2 years ago

PointsRangeFilter point_cloud_range in data_cfg is the same as in waymo-fsd.py which is [-74.88, -74.88, -1.999, 74.88, 74.88, 3.999]，and the point_cloud_range in fsd_waymoD1_1x is also differenr from that in data_cfg, How can I change it accordingly?

XLMore commented 2 years ago

PointsRangeFilter point_cloud_range in data_cfg is the same as in waymo-fsd.py which is [-74.88, -74.88, -1.999, 74.88, 74.88, 3.999]，and the point_cloud_range in fsd_waymoD1_1x is also differenr from that in data_cfg, How can I change it accordingly?

point_cloud_range in data_cfg must be a sub_set of point_cloud_range outside,I'll give it a try

XLMore commented 2 years ago

what is the small margin in z axis of point_cloud_range of data_cfg used for?

Abyssaledge commented 2 years ago

You are supposed to change to point_cloud_range in data config to fit your dataset. The small margin makes the voxelization more safe.

XLMore commented 2 years ago

I changed the point_cloud_range according to your advice and get the train loop run. But after some iterations,the loss curve looks very strange image image Is it normal because it's still in the early epoches? I tested one checkpoint and the output seems just not converged, but it's still very stange image

XLMore commented 2 years ago

I think the sparse_shape in VoteSegmentor backnone is also affected by the point cloud range, but I have no clue, Do you have any advice?

Abyssaledge commented 2 years ago

Note that we use EnableFSDDetectionHookIter hook, which means FSD model will first train segmentor for couple of iterations for warmup and then enable following detection part. So the early epochs may not effective.