Tutorial 2: Customize Datasets

Support new data format

To support a new data format, you can either convert them to existing formats or directly convert them to the middle format. You could also choose to convert them offline (before training by a script) or online (implement a new dataset and do the conversion at training). In MMDetection3D, for the data that is inconvenient to read directly online, we recommend to convert it into KITTI format and do the conversion offline, thus you only need to modify the config’s data annotation paths and classes after the conversion. For data sharing similar format with existing datasets, like Lyft compared to nuScenes, we recommend to directly implement data converter and dataset class. During the procedure, inheritation could be taken into consideration to reduce the implementation workload.

Reorganize new data formats to existing format

For data that is inconvenient to read directly online, the simplest way is to convert your dataset to existing dataset formats.

Typically we need a data converter to reorganize the raw data and convert the annotation format into KITTI style. Then a new dataset class inherited from existing ones is sometimes necessary for dealing with some specific differences between datasets. Finally, the users need to further modify the config files to use the dataset. An example training predefined models on Waymo dataset by converting it into KITTI style can be taken for reference.

Reorganize new data format to middle format

It is also fine if you do not want to convert the annotation format to existing formats. Actually, we convert all the supported datasets into pickle files, which summarize useful information for model training and inference.

The annotation of a dataset is a list of dict, each dict corresponds to a frame. A basic example (used in KITTI) is as follows. A frame consists of several keys, like image, point_cloud, calib and annos. As long as we could directly read data according to these information, the organization of raw data could also be different from existing ones. With this design, we provide an alternative choice for customizing datasets.


[
    {'image': {'image_idx': 0, 'image_path': 'training/image_2/000000.png', 'image_shape': array([ 370, 1224], dtype=int32)},
     'point_cloud': {'num_features': 4, 'velodyne_path': 'training/velodyne/000000.bin'},
     'calib': {'P0': array([[707.0493,   0.    , 604.0814,   0.    ],
       [  0.    , 707.0493, 180.5066,   0.    ],
       [  0.    ,   0.    ,   1.    ,   0.    ],
       [  0.    ,   0.    ,   0.    ,   1.    ]]),
       'P1': array([[ 707.0493,    0.    ,  604.0814, -379.7842],
       [   0.    ,  707.0493,  180.5066,    0.    ],
       [   0.    ,    0.    ,    1.    ,    0.    ],
       [   0.    ,    0.    ,    0.    ,    1.    ]]),
       'P2': array([[ 7.070493e+02,  0.000000e+00,  6.040814e+02,  4.575831e+01],
       [ 0.000000e+00,  7.070493e+02,  1.805066e+02, -3.454157e-01],
       [ 0.000000e+00,  0.000000e+00,  1.000000e+00,  4.981016e-03],
       [ 0.000000e+00,  0.000000e+00,  0.000000e+00,  1.000000e+00]]),
       'P3': array([[ 7.070493e+02,  0.000000e+00,  6.040814e+02, -3.341081e+02],
       [ 0.000000e+00,  7.070493e+02,  1.805066e+02,  2.330660e+00],
       [ 0.000000e+00,  0.000000e+00,  1.000000e+00,  3.201153e-03],
       [ 0.000000e+00,  0.000000e+00,  0.000000e+00,  1.000000e+00]]),
       'R0_rect': array([[ 0.9999128 ,  0.01009263, -0.00851193,  0.        ],
       [-0.01012729,  0.9999406 , -0.00403767,  0.        ],
       [ 0.00847068,  0.00412352,  0.9999556 ,  0.        ],
       [ 0.        ,  0.        ,  0.        ,  1.        ]]),
       'Tr_velo_to_cam': array([[ 0.00692796, -0.9999722 , -0.00275783, -0.02457729],
       [-0.00116298,  0.00274984, -0.9999955 , -0.06127237],
       [ 0.9999753 ,  0.00693114, -0.0011439 , -0.3321029 ],
       [ 0.        ,  0.        ,  0.        ,  1.        ]]),
       'Tr_imu_to_velo': array([[ 9.999976e-01,  7.553071e-04, -2.035826e-03, -8.086759e-01],
       [-7.854027e-04,  9.998898e-01, -1.482298e-02,  3.195559e-01],
       [ 2.024406e-03,  1.482454e-02,  9.998881e-01, -7.997231e-01],
       [ 0.000000e+00,  0.000000e+00,  0.000000e+00,  1.000000e+00]])},
     'annos': {'name': array(['Pedestrian'], dtype='<U10'), 'truncated': array([0.]), 'occluded': array([0]), 'alpha': array([-0.2]), 'bbox': array([[712.4 , 143.  , 810.73, 307.92]]), 'dimensions': array([[1.2 , 1.89, 0.48]]), 'location': array([[1.84, 1.47, 8.41]]), 'rotation_y': array([0.01]), 'score': array([0.]), 'index': array([0], dtype=int32), 'group_ids': array([0], dtype=int32), 'difficulty': array([0], dtype=int32), 'num_points_in_gt': array([377], dtype=int32)}}
    ...
]

On top of this you can write a new Dataset class inherited from Custom3DDataset, and overwrite related methods, like KittiDataset and ScanNetDataset.

An example of customized dataset

Here we provide an example of customized dataset.

Assume the annotation has been reorganized into a list of dict in pickle files like ScanNet. The bounding boxes annotations are stored in annotation.pkl as the following

{'point_cloud': {'num_features': 6, 'lidar_idx': 'scene0000_00'}, 'pts_path': 'points/scene0000_00.bin',
 'pts_instance_mask_path': 'instance_mask/scene0000_00.bin', 'pts_semantic_mask_path': 'semantic_mask/scene0000_00.bin',
 'annos': {'gt_num': 27, 'name': array(['window', 'window', 'table', 'counter', 'curtain', 'curtain',
       'desk', 'cabinet', 'sink', 'garbagebin', 'garbagebin',
       'garbagebin', 'sofa', 'refrigerator', 'table', 'table', 'toilet',
       'bed', 'cabinet', 'cabinet', 'cabinet', 'cabinet', 'cabinet',
       'cabinet', 'door', 'door', 'door'], dtype='<U12'),
       'location': array([[ 1.48129511,  3.52074146,  1.85652947],
       [ 2.90395617, -3.48033905,  1.52682471]]),
       'dimensions': array([[1.74445975, 0.23195696, 0.57235193],
       [0.66077662, 0.17072392, 0.67153597]]),
       'gt_boxes_upright_depth': array([
       [ 1.48129511,  3.52074146,  1.85652947,  1.74445975,  0.23195696,
         0.57235193],
       [ 2.90395617, -3.48033905,  1.52682471,  0.66077662,  0.17072392,
         0.67153597]]),
       'index': array([ 0,  1 ], dtype=int32),
       'class': array([ 6,  6 ])}}

We can create a new dataset in mmdet3d/datasets/my_dataset.py to load the data.

import numpy as np
from os import path as osp

from mmdet3d.core import show_result
from mmdet3d.core.bbox import DepthInstance3DBoxes
from mmdet.datasets import DATASETS
from .custom_3d import Custom3DDataset


@DATASETS.register_module()
class MyDataset(Custom3DDataset):
    CLASSES = ('cabinet', 'bed', 'chair', 'sofa', 'table', 'door', 'window',
               'bookshelf', 'picture', 'counter', 'desk', 'curtain',
               'refrigerator', 'showercurtrain', 'toilet', 'sink', 'bathtub',
               'garbagebin')

    def __init__(self,
                 data_root,
                 ann_file,
                 pipeline=None,
                 classes=None,
                 modality=None,
                 box_type_3d='Depth',
                 filter_empty_gt=True,
                 test_mode=False):
        super().__init__(
            data_root=data_root,
            ann_file=ann_file,
            pipeline=pipeline,
            classes=classes,
            modality=modality,
            box_type_3d=box_type_3d,
            filter_empty_gt=filter_empty_gt,
            test_mode=test_mode)

    def get_ann_info(self, index):
        # Use index to get the annos, thus the evalhook could also use this api
        info = self.data_infos[index]
        if info['annos']['gt_num'] != 0:
            gt_bboxes_3d = info['annos']['gt_boxes_upright_depth'].astype(
                np.float32)  # k, 6
            gt_labels_3d = info['annos']['class'].astype(np.long)
        else:
            gt_bboxes_3d = np.zeros((0, 6), dtype=np.float32)
            gt_labels_3d = np.zeros((0, ), dtype=np.long)

        # to target box structure
        gt_bboxes_3d = DepthInstance3DBoxes(
            gt_bboxes_3d,
            box_dim=gt_bboxes_3d.shape[-1],
            with_yaw=False,
            origin=(0.5, 0.5, 0.5)).convert_to(self.box_mode_3d)

        pts_instance_mask_path = osp.join(self.data_root,
                                          info['pts_instance_mask_path'])
        pts_semantic_mask_path = osp.join(self.data_root,
                                          info['pts_semantic_mask_path'])

        anns_results = dict(
            gt_bboxes_3d=gt_bboxes_3d,
            gt_labels_3d=gt_labels_3d,
            pts_instance_mask_path=pts_instance_mask_path,
            pts_semantic_mask_path=pts_semantic_mask_path)
        return anns_results

Then in the config, to use MyDataset you can modify the config as the following

dataset_A_train = dict(
    type='MyDataset',
    ann_file = 'annotation.pkl',
    pipeline=train_pipeline
)

Customize datasets by dataset wrappers

MMDetection3D also supports many dataset wrappers to mix the dataset or modify the dataset distribution for training like MMDetection. Currently it supports to three dataset wrappers as below:

  • RepeatDataset: simply repeat the whole dataset.

  • ClassBalancedDataset: repeat dataset in a class balanced manner.

  • ConcatDataset: concat datasets.

Repeat dataset

We use RepeatDataset as wrapper to repeat the dataset. For example, suppose the original dataset is Dataset_A, to repeat it, the config looks like the following

dataset_A_train = dict(
        type='RepeatDataset',
        times=N,
        dataset=dict(  # This is the original config of Dataset_A
            type='Dataset_A',
            ...
            pipeline=train_pipeline
        )
    )

Class balanced dataset

We use ClassBalancedDataset as wrapper to repeat the dataset based on category frequency. The dataset to repeat needs to instantiate function self.get_cat_ids(idx) to support ClassBalancedDataset. For example, to repeat Dataset_A with oversample_thr=1e-3, the config looks like the following

dataset_A_train = dict(
        type='ClassBalancedDataset',
        oversample_thr=1e-3,
        dataset=dict(  # This is the original config of Dataset_A
            type='Dataset_A',
            ...
            pipeline=train_pipeline
        )
    )

You may refer to source code for details.

Concatenate dataset

There are three ways to concatenate the dataset.

  1. If the datasets you want to concatenate are in the same type with different annotation files, you can concatenate the dataset configs like the following.

    dataset_A_train = dict(
        type='Dataset_A',
        ann_file = ['anno_file_1', 'anno_file_2'],
        pipeline=train_pipeline
    )
    

    If the concatenated dataset is used for test or evaluation, this manner supports to evaluate each dataset separately. To test the concatenated datasets as a whole, you can set separate_eval=False as below.

    dataset_A_train = dict(
        type='Dataset_A',
        ann_file = ['anno_file_1', 'anno_file_2'],
        separate_eval=False,
        pipeline=train_pipeline
    )
    
  2. In case the dataset you want to concatenate is different, you can concatenate the dataset configs like the following.

    dataset_A_train = dict()
    dataset_B_train = dict()
    
    data = dict(
        imgs_per_gpu=2,
        workers_per_gpu=2,
        train = [
            dataset_A_train,
            dataset_B_train
        ],
        val = dataset_A_val,
        test = dataset_A_test
        )
    

    If the concatenated dataset is used for test or evaluation, this manner also supports to evaluate each dataset separately.

  3. We also support to define ConcatDataset explicitly as the following.

    dataset_A_val = dict()
    dataset_B_val = dict()
    
    data = dict(
        imgs_per_gpu=2,
        workers_per_gpu=2,
        train=dataset_A_train,
        val=dict(
            type='ConcatDataset',
            datasets=[dataset_A_val, dataset_B_val],
            separate_eval=False))
    

    This manner allows users to evaluate all the datasets as a single one by setting separate_eval=False.

Note:

  1. The option separate_eval=False assumes the datasets use self.data_infos during evaluation. Therefore, COCO datasets do not support this behavior since COCO datasets do not fully rely on self.data_infos for evaluation. Combining different types of datasets and evaluating them as a whole is not tested thus is not suggested.

  2. Evaluating ClassBalancedDataset and RepeatDataset is not supported thus evaluating concatenated datasets of these types is also not supported.

A more complex example that repeats Dataset_A and Dataset_B by N and M times, respectively, and then concatenates the repeated datasets is as the following.

dataset_A_train = dict(
    type='RepeatDataset',
    times=N,
    dataset=dict(
        type='Dataset_A',
        ...
        pipeline=train_pipeline
    )
)
dataset_A_val = dict(
    ...
    pipeline=test_pipeline
)
dataset_A_test = dict(
    ...
    pipeline=test_pipeline
)
dataset_B_train = dict(
    type='RepeatDataset',
    times=M,
    dataset=dict(
        type='Dataset_B',
        ...
        pipeline=train_pipeline
    )
)
data = dict(
    imgs_per_gpu=2,
    workers_per_gpu=2,
    train = [
        dataset_A_train,
        dataset_B_train
    ],
    val = dataset_A_val,
    test = dataset_A_test
)

Modify Dataset Classes

With existing dataset types, we can modify the class names of them to train subset of the annotations. For example, if you want to train only three classes of the current dataset, you can modify the classes of dataset. The dataset will filter out the ground truth boxes of other classes automatically.

classes = ('person', 'bicycle', 'car')
data = dict(
    train=dict(classes=classes),
    val=dict(classes=classes),
    test=dict(classes=classes))

MMDetection V2.0 also supports to read the classes from a file, which is common in real applications. For example, assume the classes.txt contains the name of classes as the following.

person
bicycle
car

Users can set the classes as a file path, the dataset will load it and convert it to a list automatically.

classes = 'path/to/classes.txt'
data = dict(
    train=dict(classes=classes),
    val=dict(classes=classes),
    test=dict(classes=classes))

Note (related to MMDetection):

  • Before MMDetection v2.5.0, the dataset will filter out the empty GT images automatically if the classes are set and there is no way to disable that through config. This is an undesirable behavior and introduces confusion because if the classes are not set, the dataset only filter the empty GT images when filter_empty_gt=True and test_mode=False. After MMDetection v2.5.0, we decouple the image filtering process and the classes modification, i.e., the dataset will only filter empty GT images when filter_empty_gt=True and test_mode=False, no matter whether the classes are set. Thus, setting the classes only influences the annotations of classes used for training and users could decide whether to filter empty GT images by themselves.

  • Since the middle format only has box labels and does not contain the class names, when using CustomDataset, users cannot filter out the empty GT images through configs but only do this offline.

  • The features for setting dataset classes and dataset filtering will be refactored to be more user-friendly in the future (depends on the progress).