ADLab-AutoDrive / BEVFusion

Offical PyTorch implementation of "BEVFusion: A Simple and Robust LiDAR-Camera Fusion Framework"
Apache License 2.0
754 stars 102 forks source link

Question about nuscenes validation performance #74

Open zoomkinseok opened 1 year ago

zoomkinseok commented 1 year ago

Thank you for your exciting research. I evaluated it with the model weights (nuscenes detection validation, BEVfusion*) you provided. It came out the same as the performance reported in the paper. (mAP: 69.6, NDS: 72.1)

The command is as follows. ./tools/dist_test.sh configs/bevfusion/bevf_tf_4x8_10e_nusc_aug.py weights/bevfusion_tf.pth 8 --eval bbox

But is this result without test time augmentation applied? I ask because it seems that MultiScaleFlipAug3D is applied to the test pipeline.

Best, Minseok Joo

liyih commented 11 months ago

@zoomkinseok Is there any difference of performance between with or without MultiScaleFlipAug3D in test pipeline? Best