Filter Data through Your Query#

Jupyter Notebook

In this notebook example, we’ll take a look at Datumaro filter API.

[1]:

# Copyright (C) 2022 Intel Corporation
#
# SPDX-License-Identifier: MIT

import datumaro as dm

Filtered by subset#

We export sample VOC dataset to filter only train subset.

[2]:

dataset = dm.Dataset.import_from("./tests/assets/voc_dataset/voc_dataset1", format="voc")

[3]:

print("Representation for sample VOC dataset")
dataset

Representation for sample VOC dataset

[3]:

Dataset
        size=2
        source_path=./tests/assets/voc_dataset/voc_dataset1
        media_type=<class 'datumaro.components.media.Image'>
        annotated_items_count=1
        annotations_count=15
subsets
        test: # of items=1, # of annotated items=0, # of annotations=0, annotation types=[]
        train: # of items=1, # of annotated items=1, # of annotations=15, annotation types=['label', 'bbox', 'mask']
categories
        label: ['background', 'aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car', 'cat', 'chair', 'cow', 'diningtable', 'dog', 'horse', 'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor', 'ignored', 'head', 'hand', 'foot']
        mask: []

In VOC dataset, there are ‘train’ and ‘test’ subset. We will filter only ‘train’ subset.

[4]:

dataset = dm.Dataset.filter(dataset, '/item[subset="train"]')

[5]:

print("Representation for `train` subset of sample VOC dataset")
dataset

Representation for `train` subset of sample VOC dataset

[5]:

Dataset
        size=1
        source_path=./tests/assets/voc_dataset/voc_dataset1
        media_type=<class 'datumaro.components.media.Image'>
        annotated_items_count=1
        annotations_count=15
subsets
        train: # of items=1, # of annotated items=1, # of annotations=15, annotation types=['label', 'bbox', 'mask']
categories
        label: ['background', 'aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car', 'cat', 'chair', 'cow', 'diningtable', 'dog', 'horse', 'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor', 'ignored', 'head', 'hand', 'foot']
        mask: []

Filtered by id#

We export sample widerface dataset to filter only dataset which id is id=0_Parade_image_01.

[6]:

dataset = dm.Dataset.import_from("./tests/assets/widerface_dataset")

[7]:

print("Representation for sample WiderFace dataset")
dataset

Representation for sample WiderFace dataset

[7]:

Dataset
        size=3
        source_path=./tests/assets/widerface_dataset
        media_type=<class 'datumaro.components.media.Image'>
        annotated_items_count=3
        annotations_count=9
subsets
        train: # of items=2, # of annotated items=2, # of annotations=5, annotation types=['label', 'bbox']
        val: # of items=1, # of annotated items=1, # of annotations=4, annotation types=['label', 'bbox']
categories
        label: ['Parade', 'Handshaking']

[8]:

dataset = dm.Dataset.filter(dataset, '/item[id="0_Parade_image_01"]')

[9]:

print("Representation for `id == 1` dataset of sample WiderFace dataset")
dataset

Representation for `id == 1` dataset of sample WiderFace dataset

[9]:

Dataset
        size=1
        source_path=./tests/assets/widerface_dataset
        media_type=<class 'datumaro.components.media.Image'>
        annotated_items_count=1
        annotations_count=2
subsets
        train: # of items=1, # of annotated items=1, # of annotations=2, annotation types=['label', 'bbox']
categories
        label: ['Parade', 'Handshaking']

Filtered by width and height#

We export sample dataset to extract a dataset with images which have shorter width than height.

[10]:

dataset = dm.Dataset.import_from("./tests/assets/coco_dataset/coco")

WARNING:root:Not implemented: Found potentially conflicting source types with labels: labels, panoptic, stuff, person_keypoints, instances. Only one type will be used: instances
WARNING:root:Not implemented: conflicting source './tests/assets/coco_dataset/coco/annotations/labels_train.json' is skipped.
WARNING:root:Not implemented: conflicting source './tests/assets/coco_dataset/coco/annotations/person_keypoints_train.json' is skipped.
WARNING:root:Not implemented: conflicting source './tests/assets/coco_dataset/coco/annotations/stuff_train.json' is skipped.
WARNING:root:Not implemented: conflicting source './tests/assets/coco_dataset/coco/annotations/panoptic_train.json' is skipped.
WARNING:root:Not implemented: conflicting source './tests/assets/coco_dataset/coco/annotations/panoptic_val.json' is skipped.
WARNING:root:Not implemented: conflicting source './tests/assets/coco_dataset/coco/annotations/labels_val.json' is skipped.
WARNING:root:Not implemented: conflicting source './tests/assets/coco_dataset/coco/annotations/person_keypoints_val.json' is skipped.
WARNING:root:Not implemented: conflicting source './tests/assets/coco_dataset/coco/annotations/stuff_val.json' is skipped.

[11]:

def get_width_height(dataset: dm.Dataset):
    size_dict = {}
    for item in dataset:
        size_dict[item.id] = item.media.size
    return size_dict

[12]:

print("Representation for sample COCO dataset")
dataset

Representation for sample COCO dataset

[12]:

Dataset
        size=2
        source_path=./tests/assets/coco_dataset/coco
        media_type=<class 'datumaro.components.media.Image'>
        annotated_items_count=2
        annotations_count=6
subsets
        train: # of items=1, # of annotated items=1, # of annotations=2, annotation types=['bbox', 'caption']
        val: # of items=1, # of annotated items=1, # of annotations=4, annotation types=['mask', 'caption', 'polygon']
categories
        label: ['a', 'b', 'c']

[13]:

print("Width and height for sample COCO dataset images")
get_width_height(dataset)

Width and height for sample COCO dataset images

[13]:

{'a': (5, 10), 'b': (10, 5)}

[14]:

dataset = dm.Dataset.filter(dataset, "/item[image/width < image/height]")

[15]:

print("Representation for `width < height` sample COCO dataset images")
dataset

Representation for `width < height` sample COCO dataset images

[15]:

Dataset
        size=1
        source_path=./tests/assets/coco_dataset/coco
        media_type=<class 'datumaro.components.media.Image'>
        annotated_items_count=1
        annotations_count=4
subsets
        val: # of items=1, # of annotated items=1, # of annotations=4, annotation types=['mask', 'caption', 'polygon']
categories
        label: ['a', 'b', 'c']

[16]:

print("Width and height for `width < height` sample COCO dataset images")
get_width_height(dataset)

Width and height for `width < height` sample COCO dataset images

[16]:

{'b': (10, 5)}

Filtered by label and area#

We export sample dataset to extract only non-persons.

[17]:

dataset = dm.Dataset.import_from("./tests/assets/voc_dataset/voc_dataset1")

[18]:

print("Representation for sample VOC dataset")
dataset

Representation for sample VOC dataset

[18]:

Dataset
        size=2
        source_path=./tests/assets/voc_dataset/voc_dataset1
        media_type=<class 'datumaro.components.media.Image'>
        annotated_items_count=1
        annotations_count=15
subsets
        test: # of items=1, # of annotated items=0, # of annotations=0, annotation types=[]
        train: # of items=1, # of annotated items=1, # of annotations=15, annotation types=['label', 'bbox', 'mask']
categories
        label: ['background', 'aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car', 'cat', 'chair', 'cow', 'diningtable', 'dog', 'horse', 'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor', 'ignored', 'head', 'hand', 'foot']
        mask: []

Indicate filter_annotations as True if filter needs to apply to annotations. The default value is False to items.

[19]:

dataset = dm.Dataset.filter(dataset, '/item/annotation[label!="person"]', filter_annotations=True)

[20]:

print('Representation for sample VOC dataset whose annotation is `label!="person"`')
dataset

Representation for sample VOC dataset whose annotation is `label!="person"`

[20]:

Dataset
        size=2
        source_path=./tests/assets/voc_dataset/voc_dataset1
        media_type=<class 'datumaro.components.media.Image'>
        annotated_items_count=1
        annotations_count=13
subsets
        test: # of items=1, # of annotated items=0, # of annotations=0, annotation types=[]
        train: # of items=1, # of annotated items=1, # of annotations=13, annotation types=['label', 'bbox', 'mask']
categories
        label: ['background', 'aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car', 'cat', 'chair', 'cow', 'diningtable', 'dog', 'horse', 'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor', 'ignored', 'head', 'hand', 'foot']
        mask: []

Filtered by annotation#

We export sample dataset to extract non-occluded annotations, remove empty images. Use data only from the “s1” source of the project.

[21]:

dataset = dm.Dataset.import_from("./tests/assets/voc_dataset/voc_dataset1")

[22]:

print("Representation for sample VOC dataset")
dataset

Representation for sample VOC dataset

[22]:

Dataset
        size=2
        source_path=./tests/assets/voc_dataset/voc_dataset1
        media_type=<class 'datumaro.components.media.Image'>
        annotated_items_count=1
        annotations_count=15
subsets
        test: # of items=1, # of annotated items=0, # of annotations=0, annotation types=[]
        train: # of items=1, # of annotated items=1, # of annotations=15, annotation types=['label', 'bbox', 'mask']
categories
        label: ['background', 'aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car', 'cat', 'chair', 'cow', 'diningtable', 'dog', 'horse', 'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor', 'ignored', 'head', 'hand', 'foot']
        mask: []

[23]:

dm.Dataset.filter(
    dataset, '/item/annotation[occluded="False"]', filter_annotations=True, remove_empty=True
)

[23]:

Dataset
        size=1
        source_path=./tests/assets/voc_dataset/voc_dataset1
        media_type=<class 'datumaro.components.media.Image'>
        annotated_items_count=1
        annotations_count=2
subsets
        train: # of items=1, # of annotated items=1, # of annotations=2, annotation types=['bbox']
categories
        label: ['background', 'aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car', 'cat', 'chair', 'cow', 'diningtable', 'dog', 'horse', 'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor', 'ignored', 'head', 'hand', 'foot']
        mask: []

[24]:

print("Representation for `non-occluded annotations and empty images removed sample VOC dataset`")
dataset

Representation for `non-occluded annotations and empty images removed sample VOC dataset`

[24]:

Dataset
        size=1
        source_path=./tests/assets/voc_dataset/voc_dataset1
        media_type=<class 'datumaro.components.media.Image'>
        annotated_items_count=1
        annotations_count=2
subsets
        train: # of items=1, # of annotated items=1, # of annotations=2, annotation types=['bbox']
categories
        label: ['background', 'aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car', 'cat', 'chair', 'cow', 'diningtable', 'dog', 'horse', 'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor', 'ignored', 'head', 'hand', 'foot']
        mask: []