search

Tsinghua-MARS-Lab / futr3d

Code for paper: FUTR3D: a unified sensor fusion framework for 3d detection
Apache License 2.0
282 stars 39 forks source link

lidar_0075_900q.pth is actually for 600 queries #29

Open mrvangeerenstein opened 1 year ago

mrvangeerenstein commented 1 year ago

Hi there,

Thanks a lot for releasing the new version of your project!

I have a small note regarding the available checkpoint for the LiDAR-only model. The checkpoint and corresponding cfg file mention the use of num_queries=900, but when loading the checkpoint it appears that it is actually trained with 600 queries;

size mismatch for pts_bbox_head.tgt_embed.weight: copying a param with shape torch.Size([600, 256]) from checkpoint, the shape in current model is torch.Size([900, 256]). size mismatch for pts_bbox_head.refpoint_embed.weight: copying a param with shape torch.Size([600, 3]) from checkpoint, the shape in current model is torch.Size([900, 3]).

For the LiDAR+camera checkpoints I don't see anything like this, so these are indeed with 900 queries.

Will you release the checkpoint for the LiDAR-only model with 900 queries? I assume this yields different results than a model with 600 queries.

Thank you!

xyaochen commented 1 year ago

Thanks for pointing this out! I have corrected the LiDAR checkpoint link