Abnormal camera branch training values

[artofstate]

Hi！When I run the code you provided, the results of lidar stream are as follows: mAP :0.6534, NDS:0.7022. While the mAＰis only 0.0127 and NDS is 0.0420. The training log shows that after 20 epochs, the loss is abuout 5.8, and the mached_ious is only about 0.14. That's really strange!
I performed the validation on the nuscenes dataset. Have you ever encounted the same issue?