search

cxy1997 / 3D_adapt_auto_driving

Train in Germany, Test in The USA: Making 3D Object Detectors Generalize
https://arxiv.org/abs/2005.08139
130 stars 13 forks source link

About nuScenes #7

Open guangyaooo opened 3 years ago

guangyaooo commented 3 years ago

Hello, Thanks for the great code,it is very helpful to me. I want to reproduce the nuscenes results in the paper, but it seems that the KittiRCNNDataset is incompatible with train_rcnn.py, https://github.com/cxy1997/3D_adapt_auto_driving/blob/72389c25490caba06b8e06b80452e7d7c5e7b241/pointrcnn/lib/datasets/kitti_rcnn_dataset.py#L12-L15 https://github.com/cxy1997/3D_adapt_auto_driving/blob/72389c25490caba06b8e06b80452e7d7c5e7b241/pointrcnn/tools/train_rcnn.py#L72-L78 the __init__ function of KittiRCNNDataset has no parameters named npoints_faraway, with_replace, subsample and shuffle_subsample, so I directly dropped these parameters in train_rcnn. py.

Then I trained pointrcnn with default.yaml on kitti-format nuscens dataset without any rescaled data. The results are shown in the figure below

image with old metirc

image with new metric

The results reported in the paper: image

As we can see, The performance of both the new and old metrics is much lower than that reported in the paper. So could you please provide more details about the nuScenes training, such as nuScenes training config.

mileyan commented 3 years ago

Hi @Dgy2017, the pointRCNN version is wrong. I will update the correct code soon. Thanks.