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

Question about performance of pretrained model #88

Closed yuchen-he closed 1 year ago

yuchen-he commented 1 year ago

I have tried to pretrain a SpConv based segmentation model on my custom dataset. But it gave poor segmentation performance when inferring a single frame point cloud. Have you ever tried to evaluate the performance of segmentation model only? or do you have any idea about the poor performance?

Here is my training config.

fsd_pretrain_local.txt

Abyssaledge commented 1 year ago

I have not validated the performance of the segmentation part. It's not recommended to judge the performance by a single-frame result.

yuchen-he commented 1 year ago

I have not validated the performance of the segmentation part. It's not recommended to judge the performance by a single-frame result.

Thanks a lot. I will close this issue.