Tutorial 2: Customize Datasets

Support new data format

To support a new data format, you can either convert them to existing formats (COCO format or PASCAL format) 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 MMDetection, we recommand to convert the data into COCO formats and do the conversion offline, thus you only need to modify the config’s data annotation pathes and classes after the conversion to your data.

Reorganize new data formats to existing format

The simplest way is to convert your dataset to existing dataset formats (COCO or PASCAL VOC).

The annotation json files in COCO format has the following necessary keys:

'images': [
    {
        'file_name': 'COCO_val2014_000000001268.jpg',
        'height': 427,
        'width': 640,
        'id': 1268
    },
    ...
],

'annotations': [
    {
        'segmentation': [[192.81,
            247.09,
            ...
            219.03,
            249.06]],  # if you have mask labels
        'area': 1035.749,
        'iscrowd': 0,
        'image_id': 1268,
        'bbox': [192.81, 224.8, 74.73, 33.43],
        'category_id': 16,
        'id': 42986
    },
    ...
],

'categories': [
    {'id': 0, 'name': 'car'},
 ]

There are three necessary keys in the json file:

  • images: contains a list of images with their informations like file_name, height, width, and id.
  • annotations: contains the list of instance annotations.
  • categories: contains the list of categories names and their ID.

After the data pre-processing, the users need to further modify the config files to use the dataset. Here we show an example of using a custom dataset of 5 classes, assuming it is also in COCO format.

In configs/my_custom_config.py:

...
# dataset settings
dataset_type = 'CocoDataset'
classes = ('a', 'b', 'c', 'd', 'e')
...
data = dict(
    samples_per_gpu=2,
    workers_per_gpu=2,
    train=dict(
        type=dataset_type,
        classes=classes,
        ann_file='path/to/your/train/data',
        ...),
    val=dict(
        type=dataset_type,
        classes=classes,
        ann_file='path/to/your/val/data',
        ...),
    test=dict(
        type=dataset_type,
        classes=classes,
        ann_file='path/to/your/test/data',
        ...))
...

We use this way to support CityScapes dataset. The script is in cityscapes.py and we also provide the finetuning configs.

Note

  1. For instance segmentation datasets, MMDetection only supports evaluating mask AP of dataset in COCO format for now.
  2. It is recommanded to convert the data offline before training, thus you can still use CocoDataset and only need to modify the path of annotations and the training classes.

Reorganize new data format to middle format

It is also fine if you do not want to convert the annotation format to COCO or PASCAL format. Actually, we define a simple annotation format and all existing datasets are processed to be compatible with it, either online or offline.

The annotation of a dataset is a list of dict, each dict corresponds to an image. There are 3 field filename (relative path), width, height for testing, and an additional field ann for training. ann is also a dict containing at least 2 fields: bboxes and labels, both of which are numpy arrays. Some datasets may provide annotations like crowd/difficult/ignored bboxes, we use bboxes_ignore and labels_ignore to cover them.

Here is an example.


[
    {
        'filename': 'a.jpg',
        'width': 1280,
        'height': 720,
        'ann': {
            'bboxes': <np.ndarray, float32> (n, 4),
            'labels': <np.ndarray, int64> (n, ),
            'bboxes_ignore': <np.ndarray, float32> (k, 4),
            'labels_ignore': <np.ndarray, int64> (k, ) (optional field)
        }
    },
    ...
]

There are two ways to work with custom datasets.

  • online conversion

    You can write a new Dataset class inherited from CustomDataset, and overwrite two methods load_annotations(self, ann_file) and get_ann_info(self, idx), like CocoDataset and VOCDataset.

  • offline conversion

    You can convert the annotation format to the expected format above and save it to a pickle or json file, like pascal_voc.py. Then you can simply use CustomDataset.

An example of customized dataset

Assume the annotation is in a new format in text files. The bounding boxes annotations are stored in text file annotation.txt as the following

#
000001.jpg
1280 720
2
10 20 40 60 1
20 40 50 60 2
#
000002.jpg
1280 720
3
50 20 40 60 2
20 40 30 45 2
30 40 50 60 3

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

import mmcv
import numpy as np

from .builder import DATASETS
from .custom import CustomDataset


@DATASETS.register_module()
class MyDataset(CustomDataset):

    CLASSES = ('person', 'bicycle', 'car', 'motorcycle')

    def load_annotations(self, ann_file):
        ann_list = mmcv.list_from_file(ann_file)

        data_infos = []
        for i, ann_line in enumerate(ann_list):
            if ann_line != '#':
                continue

            img_shape = ann_list[i + 2].split(' ')
            width = int(img_shape[0])
            height = int(img_shape[1])
            bbox_number = int(ann_list[i + 3])

            anns = ann_line.split(' ')
            bboxes = []
            labels = []
            for anns in ann_list[i + 4:i + 4 + bbox_number]:
                bboxes.append([float(ann) for ann in anns[:4]])
                labels.append(int(anns[4]))

            data_infos.append(
                dict(
                    filename=ann_list[i + 1],
                    width=width,
                    height=height,
                    ann=dict(
                        bboxes=np.array(bboxes).astype(np.float32),
                        labels=np.array(labels).astype(np.int64))
                ))

        return data_infos

    def get_ann_info(self, idx):
        return self.data_infos[idx]['ann']

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

dataset_A_train = dict(
    type='MyDataset',
    ann_file = 'image_list.txt',
    pipeline=train_pipeline
)

Customize datasets by dataset wrappers

MMDetection also supports many dataset wrappers to mix the dataset or modify the dataset distribution for training. 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:

  • 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 v2.8.0 or v2.9.0 (depends on the progress).