The version of mmdet3d about sparse4d HOT 18 CLOSED

Our code does not rely much on mmdet3d, mainly the box3d structure, but the structure of mmdet3d has not been modified for a long time, so the latest version of mmdet3d is also supported.
You can try the inference test first. If the metrics are consistent with the published ones, it means that there is no problem.

from sparse4d.

Thank you. By the way, where is the file "nuscenes_infos_trainval_with_inds.pkl" used? I can't find another usage except for README.md.

from sparse4d.

This is for multi-object tracking task, but we haven't released the tracking code yet.

from sparse4d.

This is for multi-object tracking task, but we haven't released the tracking code yet.

Thanks for your awesome job! Do you have any release plan about tracking task?

from sparse4d.

Currently there are no plans.

from sparse4d.

@linxuewu , Hello, what is the purpose of switching L and W at: https://github.com/linxuewu/Sparse4D/blob/main/projects/mmdet3d_plugin/datasets/nuscenes_3d_det_track_dataset.py#L948

from sparse4d.

Because the order defined by mmdet3d is different from that of nuscenes.
https://github.com/open-mmlab/mmdetection3d/blob/master/tools/data_converter/nuscenes_converter.py#L256

from sparse4d.

But in mmdet3d==0.17.2, there is no such switching
https://github.com/open-mmlab/mmdetection3d/blob/v0.17.2/tools/data_converter/nuscenes_converter.py#L253

So this means that, you preprocess the nuscenes dataset through the latest mmdet3d but train/evaluate the method in mmdet3d==0.17.2 ?

from sparse4d.

It seems so. I didn't pay attention to the mmdet3d version used for data processing before...

from sparse4d.

However, training and evaluation do not rely too much on mmdet3d, so version misalignment will not cause other bugs.

from sparse4d.

It seems that the ground truth of rotation is also different in different mmdet3d versions. So the mAOE seems to be affected too.

from sparse4d.

This should be the cause of your mAOE problem, just use the new version of mmdet3d for data preprocess.

from sparse4d.

Or directly use the pkl file I provided. https://github.com/linxuewu/Sparse4D/releases/download/v0.0/nuscenes_infos_trainval_with_inds.pkl

from sparse4d.

OK, but how to use this pkl file? Simply replace "nuscenes_infos_val.pkl" with "nuscenes_infos_trainval_with_inds.pkl" in the config file as shown in the following ?

data = dict(
    samples_per_gpu=1,
    workers_per_gpu=4,
    train=dict(
        **data_basic_config,
        ann_file=anno_root + 'nuscenes_infos_train.pkl',
        pipeline=train_pipeline,
        test_mode=False,
    ),
    val=dict(
        **data_basic_config,
        ann_file=anno_root + 'nuscenes_infos_val.pkl',
        pipeline=test_pipeline,
        test_mode=True,
    ),
    test=dict(
        **data_basic_config,
        ann_file=anno_root + 'nuscenes_infos_val.pkl',
        pipeline=test_pipeline,
        test_mode=True,
    ),
)

from sparse4d.

This is the union of the training set and the validation set, prepared for test-model training. You can split it with nuscenes.utils.split.

from sparse4d.

https://github.com/linxuewu/Sparse4D/releases/download/v0.0/nuscenes_infos_train.pkl
If you only want to use train-set, use this pkl.

from sparse4d.

https://github.com/linxuewu/Sparse4D/releases/download/v0.0/nuscenes_infos_train.pkl
If you only want to use train-set, use this pkl.

Thank you for your attention. But it seems not elegant, I will submit a PR to fix the misalignment of the version of mmdetection3d in preprocessing data and training/evaluation in code.

from sparse4d.

I have submitted the PR. The problem mainly lies in your modification of "output_to_nusc_box" to fit the misalignment when you are conducting experiments with mmdet3d==0.17.2 but with the dataset preprocessed by mmdet3d>1.0.0.

#10 (comment)

from sparse4d.