Questions on Training Uni3D

[wangyubin04]

Hello！ I wonder if I want to reproduce the experimental results in Table 7 in Uni3D, how can I adjust the number of samples in nuScenes dataset, and where is the script for training on one dataset only?

[BOBrown]

@wangyubin04 Hi, experimental results in Table 7 show merging all training data from KITTI and few-shot (e.g. 5%) data from nusc. You could generate 5% nusc info path, and change the INFO_PATH in da_nuscenes_dataset_gt_D_p.yaml. One example is shown in semi-superivsed setting in our repo ada_05_nuscenes_dataset_gt_D.yaml

[wangyubin04]

Can I just directly modify the include_nuscenes_data function in nuscenes_dataset.py instead, since sampling for the waymo dataset has been similarly performed in waymo_dataset.py. Besides, I wonder how to train on only one dataset (e.g. nuscenes) in your code repository to reproduce results in Table 7.

[wangyubin04]

Another question, is the final result on nuscenes referring to the results highlighted in red boxes? I'm not sure if the location of the results is correct.

[BOBrown]

@wangyubin04 If you are reporting $AP_{3D}$ over 40 recall positions, the red box indicates the correct value for Car, Pede, Cyclist.

[BOBrown]

Can I just directly modify the include_nuscenes_data function in nuscenes_dataset.py instead, since sampling for the waymo dataset has been similarly performed in waymo_dataset.py. Besides, I wonder how to train on only one dataset (e.g. nuscenes) in your code repository to reproduce results in Table 7.

You can refer to only_nusc for reproducing the only one dataset training results.