[?] How can I use FiftyOne to convert OpenImage segmentations to COCO format

santhoshkelathodi commented 2 years ago

Proposal Summary

As there are multiple data formats for creating datasets, there can be a feature to convert one format to another. For example, OpenImages to Coco format, or coco to Yolo, etc. As I understand the visualization from different formats are supported, this can be easier.

Motivation

What is the use case for this feature?
- easy conversion of datasets across different formats to be used in different models.
Why is this use case valuable to support for FiftyOne users in general?
- While evaluating a dataset, it may be necessary to remove irrelevant annotations and then it can be converted into the relevant formats.
Why is this use case valuable to support for your project(s) or organization?
- seamless integration of dataset on different models
Why is it currently difficult to achieve this use case? (please be as specific as possible about why related FiftyOne features and components are insufficient)
- I cannot find a documentation for exporting a dataset to another type

What areas of FiftyOne does this feature affect?

[ ] App: FiftyOne application
[x] Core: Core fiftyone Python library
[ ] Server: FiftyOne server

Details

dataset = foz.load_zoo_dataset(
    "open-images-v6",
    split="test",
    label_types=["segmentations"],
    classes=["Cattle"],
    max_samples=200,
    seed=51,
    shuffle=True,
    dataset_name="open-images-cattle",
)

export_dir = "exports"
label_field = "ground_truth"  # for example
dataset_type = fo.types.COCODetectionDataset
# tagged_view below is a part of the dataset view where some edits are made. 
dataset.export(
    export_dir=export_dir,
    dataset_type=dataset_type,
    label_field=label_field,
)

Willingness to contribute

The FiftyOne Community encourages new feature contributions. Would you or another member of your organization be willing to contribute an implementation of this feature?

[ ] Yes. I can contribute this feature independently.
[x] Yes. I would be willing to contribute this feature with guidance from the FiftyOne community.
[ ] No. I cannot contribute this feature at this time.

jasoncorso commented 2 years ago

The CLI already supports a dataset format conversion command convert and it does so on disk directly. https://voxel51.com/docs/fiftyone/cli/index.html

santhoshkelathodi commented 2 years ago

The CLI already supports a dataset format conversion command convert and it does so on disk directly. https://voxel51.com/docs/fiftyone/cli/index.html

Thank you for the quick reply...

I managed to find out the conversion functions. However, I could not find a way to convert from Google OpenImage segmentation format to coco instance segmentation format. Is it that I am not able to correctly identify the format ?

https://voxel51.com/docs/fiftyone/recipes/convert_datasets.html

brimoor commented 2 years ago

@santhoshkelathodi what's the issue here? The code you shared should work, assuming you provide a valid field name for the label_field argument of export() (or just omit it altogether if your dataset only contains one Detections field, because FiftyOne can automatically infer the correct field):

For example, this works for me:

import fiftyone as fo
import fiftyone.zoo as foz

# Load Open Images dataset
dataset = foz.load_zoo_dataset(
    "open-images-v6",
    split="validation",
    label_types=["segmentations"],
    classes=["Cattle"],
    max_samples=10,
)

session = fo.launch_app(dataset)

# Export in COCO format
dataset.export(
    export_dir="/tmp/coco",
    dataset_type=fo.types.COCODetectionDataset,
    label_field="segmentations",  # this can be omitted bc dataset only contains one `Detections` field
)

# Verify that we can load the COCO dataset
dataset2 = fo.Dataset.from_dir(
    dataset_dir="/tmp/coco",
    dataset_type=fo.types.COCODetectionDataset,
    label_types="segmentations",  # required in order to load masks (otherwise only bboxes are loaded)
)

session.dataset = dataset2

brimoor commented 2 years ago

Just for fun, here's how this would look using only the CLI:

# Download subset of Open Images dataset
fiftyone zoo datasets download open-images-v6 \
    --splits validation \
    --kwargs \
        label_types=segmentations \
        classes=Cattle \
        max_samples=10

# Get location where dataset is stored
INPUT_DIR=$(fiftyone zoo datasets find open-images-v6 --split validation)

# Destination to write the COCO-formatted data
OUTPUT_DIR=/tmp/coco

# Convert segmentations from Open Images dataset to COCO format
fiftyone convert \
    --input-dir ${INPUT_DIR} \
    --input-type fiftyone.types.OpenImagesV6Dataset \
    --input-kwargs \
        label_types=segmentations \
        include_id=False \
    --output-dir ${OUTPUT_DIR} \
    --output-type fiftyone.types.COCODetectionDataset

# Check out results
ls -lah ${OUTPUT_DIR}
python -m json.tool "${OUTPUT_DIR}/labels.json"

santhoshkelathodi commented 2 years ago

Thank you very much... I am able to do the conversion...

voxel51 / fiftyone