BEVDet baseline with 42 mIOU

[cvpr2023-3d-occupancy-prediction]

BEVDet baseline with 42 mIOU

[HuangJunJie2017]

@0w0h0y Using the camera mask in calculating loss is the main reason. In fact, training BEVFormer with the camera mask in calculating loss can also improve performance to a large extend. For example, training BEVFormer-Base with the camera mask, it scores 33. 4 mIOU at epoch 4:

===> per class IoU of 6019 samples:
===> others - IoU = 5.51
===> barrier - IoU = 42.75
===> bicycle - IoU = 10.26
===> bus - IoU = 40.46
===> car - IoU = 49.09
===> construction_vehicle - IoU = 15.41
===> motorcycle - IoU = 14.86
===> pedestrian - IoU = 23.71
===> traffic_cone - IoU = 21.18
===> trailer - IoU = 26.41
===> truck - IoU = 33.04
===> driveable_surface - IoU = 79.69
===> other_flat - IoU = 35.28
===> sidewalk - IoU = 47.6
===> terrain - IoU = 50.23
===> manmade - IoU = 39.02
===> vegetation - IoU = 32.7
===> mIoU of 6019 samples: 33.37

Other than that, I haven't done a detailed ablation study on other factors between the two paradigms.

[waveleaf27]

Great work !

[0w0h0y]

I wonder what's the major factor for this strong performance, since BEVFormer shows basically flat performance with BEVDet on detection tasks.