Drawing Labels on Samples¶

This recipe demonstrates how to use FiftyOne to render annotated versions of image and video samples with their label field(s) overlaid.

Setup¶

If you haven’t already, install FiftyOne:

[ ]:

!pip install fiftyone

In this recipe we’ll use the FiftyOne Dataset Zoo to download some labeled datasets to use as sample data for drawing labels.

Behind the scenes, FiftyOne uses either the TensorFlow Datasets or TorchVision Datasets libraries to wrangle the datasets, depending on which ML library you have installed.

You can, for example, install PyTorch as follows (we’ll also need pycocotools to load the COCO dataset, in particular):

[1]:

!pip install torch torchvision
!pip install pycocotools

Drawing COCO detections¶

You can download the validation split of the COCO-2017 dataset to ~/fiftyone/coco-2017/validation by running the following command:

[1]:

!fiftyone zoo datasets download coco-2017 --splits validation

Split 'validation' already downloaded

Now let’s load the dataset, extract a DatasetView that contains 100 images from the dataset, and render them as annotated images with their ground truth labels overlaid:

[2]:

import fiftyone as fo
import fiftyone.zoo as foz
import fiftyone.utils.annotations as foua

# Directory to write the output annotations
anno_dir = "/tmp/fiftyone/draw_labels/coco-2017-validation-anno"

# Load the validation split of the COCO-2017 dataset
dataset = foz.load_zoo_dataset("coco-2017", split="validation")

# Extract some samples
view = dataset.limit(100)

#
# You can customize the look-and-feel of the annotations
# For more information, see:
# https://voxel51.com/docs/fiftyone/user_guide/draw_labels.html#customizing-label-rendering
#
config = foua.DrawConfig({
    "per_object_label_colors": True
})

# Render the labels
print("Writing annotated images to '%s'" % anno_dir)
view.draw_labels(anno_dir, config=config)
print("Annotation complete")

Split 'validation' already downloaded
Loading 'coco-2017' split 'validation'
 100% |█████| 5000/5000 [14.8s elapsed, 0s remaining, 339.4 samples/s]
Writing annotated images to '/tmp/fiftyone/draw_labels/coco-2017-validation-anno'
 100% |███████| 100/100 [7.3s elapsed, 0s remaining, 11.9 samples/s]
Annotation complete

Let’s list the output directory to verify that the annotations have been generated:

[3]:

!ls -lah /tmp/fiftyone/draw_labels/coco-2017-validation-anno | head

total 51976
drwxr-xr-x  202 Brian  wheel   6.3K Jul 27 18:36 .
drwxr-xr-x    5 Brian  wheel   160B Jul 27 15:59 ..
-rw-r--r--    1 Brian  wheel   115K Jul 27 18:36 000001-2.jpg
-rw-r--r--@   1 Brian  wheel   116K Jul 27 12:51 000001.jpg
-rw-r--r--    1 Brian  wheel   243K Jul 27 18:36 000002-2.jpg
-rw-r--r--    1 Brian  wheel   243K Jul 27 12:51 000002.jpg
-rw-r--r--    1 Brian  wheel   177K Jul 27 18:36 000003-2.jpg
-rw-r--r--@   1 Brian  wheel   177K Jul 27 12:51 000003.jpg
-rw-r--r--    1 Brian  wheel   101K Jul 27 18:36 000004-2.jpg

Here’s an example of an annotated image that was generated:

coco-2017-annotated

Drawing Caltech 101 classifications¶

You can download the test split of the Caltech 101 dataset to ~/fiftyone/caltech101/test by running the following command:

[4]:

!fiftyone zoo datasets download caltech101 --splits test

Split 'test' already downloaded

Now let’s load the dataset, extract a DatasetView that contains 100 images from the dataset, and render them as annotated images with their ground truth labels overlaid:

[5]:

import fiftyone as fo
import fiftyone.zoo as foz
import fiftyone.utils.annotations as foua

# Directory to write the output annotations
anno_dir = "/tmp/fiftyone/draw_labels/caltech101-test-anno"

# Load the test split of the Caltech 101 dataset
dataset = foz.load_zoo_dataset("caltech101", split="test")

# Extract some samples
view = dataset.limit(100)

#
# You can customize the look-and-feel of the annotations
# For more information, see:
# https://voxel51.com/docs/fiftyone/user_guide/draw_labels.html#customizing-label-rendering
#
config = foua.DrawConfig({
    "font_size": 36
})

# Render the labels
print("Writing annotated images to '%s'" % anno_dir)
view.draw_labels(anno_dir, config=config)
print("Annotation complete")

Split 'test' already downloaded
Loading 'caltech101' split 'test'
 100% |█████| 9145/9145 [4.8s elapsed, 0s remaining, 1.9K samples/s]
Writing annotated images to '/tmp/fiftyone/draw_labels/caltech101-test-anno'
 100% |███████| 100/100 [2.6s elapsed, 0s remaining, 37.4 samples/s]
Annotation complete

Let’s list the output directory to verify that the annotations have been generated:

[6]:

!ls -lah /tmp/fiftyone/draw_labels/caltech101-test-anno | head

total 17456
drwxr-xr-x  182 Brian  wheel   5.7K Jul 27 18:37 .
drwxr-xr-x    5 Brian  wheel   160B Jul 27 15:59 ..
-rw-r--r--@   1 Brian  wheel    13K Jul 27 18:37 image_0001-2.jpg
-rw-r--r--    1 Brian  wheel    41K Jul 27 15:59 image_0001.jpg
-rw-r--r--    1 Brian  wheel   197K Jul 27 18:37 image_0002.jpg
-rw-r--r--    1 Brian  wheel   5.9K Jul 27 18:37 image_0003.jpg
-rw-r--r--    1 Brian  wheel    19K Jul 27 18:37 image_0004-2.jpg
-rw-r--r--    1 Brian  wheel    33K Jul 27 15:59 image_0004.jpg
-rw-r--r--    1 Brian  wheel    18K Jul 27 18:37 image_0005-2.jpg

Here’s an example of an annotated image that was generated:

69ca7cd9a3e740ba9fc4a30c534e0a73

Drawing labels on videos¶

FiftyOne can also render frame labels onto video samples.

To demonstrate, let’s work with the (small) video quickstart dataset from the zoo:

[2]:

import fiftyone.zoo as foz

# Load a small video dataset
dataset = foz.load_zoo_dataset("quickstart-video")

print(dataset)

Dataset already downloaded
Loading 'quickstart-video'
 100% |█████████| 10/10 [15.9s elapsed, 0s remaining, 0.6 samples/s]
Name:           quickstart-video
Media type      video
Num samples:    10
Persistent:     False
Info:           {'description': 'quickstart-video'}
Tags:           []
Sample fields:
    media_type: fiftyone.core.fields.StringField
    filepath:   fiftyone.core.fields.StringField
    tags:       fiftyone.core.fields.ListField(fiftyone.core.fields.StringField)
    metadata:   fiftyone.core.fields.EmbeddedDocumentField(fiftyone.core.metadata.Metadata)
    frames:     fiftyone.core.fields.FramesField
Frame fields:
    frame_number: fiftyone.core.fields.FrameNumberField
    objs:         fiftyone.core.fields.EmbeddedDocumentField(fiftyone.core.labels.Detections)

Note that the dataset contains frame-level detections in the objs field of each frame.

Let’s make a DatasetView that contains a couple random videos from the dataset and render them as annotated videos with the frame-level detections overlaid:

[3]:

import fiftyone.utils.annotations as foua

# Directory to write the output annotations
anno_dir = "/tmp/fiftyone/draw_labels/quickstart-video-anno"

# Extract two random samples
view = dataset.take(2)

#
# You can customize the look-and-feel of the annotations
# For more information, see:
# https://voxel51.com/docs/fiftyone/user_guide/draw_labels.html#customizing-label-rendering
#
config = foua.DrawConfig({
    "per_object_label_colors": True
})

# Render the labels
print("Writing annotated videos to '%s'" % anno_dir)
view.draw_labels(anno_dir, config=config)
print("Annotation complete")

Writing annotated videos to '/tmp/fiftyone/draw_labels/quickstart-video-anno'
Rendering video 1/2: '/tmp/fiftyone/draw_labels/quickstart-video-anno/0587e1cfc2344523922652d8b227fba4-000014-video_052.mp4'
 100% |████████| 120/120 [19.0s elapsed, 0s remaining, 6.7 frames/s]
Rendering video 2/2: '/tmp/fiftyone/draw_labels/quickstart-video-anno/0587e1cfc2344523922652d8b227fba4-000014-video_164.mp4'
 100% |████████| 120/120 [27.2s elapsed, 0s remaining, 4.3 frames/s]
Annotation complete

Let’s list the output directory to verify that the annotations have been generated:

[4]:

!ls -lah /tmp/fiftyone/draw_labels/quickstart-video-anno

total 34832
drwxr-xr-x  4 Brian  wheel   128B Oct  7 23:57 .
drwxr-xr-x  3 Brian  wheel    96B Oct  7 23:57 ..
-rw-r--r--  1 Brian  wheel   7.5M Oct  7 23:57 0587e1cfc2344523922652d8b227fba4-000014-video_052.mp4
-rw-r--r--  1 Brian  wheel   8.5M Oct  7 23:58 0587e1cfc2344523922652d8b227fba4-000014-video_164.mp4

Here’s a snippet of an annotated video that was generated:

quickstart-video-annotated

Cleanup¶

You can cleanup the files generated by this recipe by running the command below:

[7]:

!rm -rf /tmp/fiftyone