Welcome to the Unity Dataset Insights documentation!¶
Unity Dataset Insights is a python package for understanding synthetic datasets. This package enables users to analyze synthetic datasets generated using the Perception SDK for the Unity game engine and, optionally, at scale using Unity Simulations. Dataset Insights exposes the metrics collected when the dataset was created e.g. object count, label distribution, etc. To use our sample notebooks pull our docker image unitytechnologies/datasetinsights
Workflow¶
Download From Unity Simulations¶
If you used Unity Simulations to generate your dataset,
then you can use download_manifest() method and the Downloader object to download your data.
manifest_file = os.path.join(data_volume, f"{run_execution_id}.csv")
download_manifest(run_execution_id, manifest_file, auth_token)
dl = Downloader(manifest_file, data_root, use_cache=True)
The first line of code defines your local path for the manifest file.
The second line downloads the manifest from Unity Simulations to the local path.
The third line, initializes the Downloader object.
Once initialized, the downloader class can download_all(),
to download all files associated with that dataset. Or use any of the following methods:
download_captures(),
download_metrics(),
to download the captures or metrics respectively.
Load Dataset Metadata¶
Once the dataset’s metadata is downloaded, it can be loaded for statistics using the simulation module.
Annotation and metric definitions are loaded into pandas dataframes using AnnotationDefinitions and MetricDefinitions respectively.
These definition ids correspond to the different kinds of metrics and annotations that are present in the dataset.
You can then use these ids to filter the statistics to correspond to the specific metric or annotation you are interested in.
For instance, for the SynthDet dataset, you can access the dataframe corresponding to foreground placement
by filtering on that id.
The code snippet below shows how you could filter the metrics to just display information about the foreground object placement and then filter again to just show the rotations of the foreground objects present in the SynthDet dataset.
foreground_placement_info_definition_id = "061e08cc-4428-4926-9933-a6732524b52b"
columns = ("x_rot", "y_rot", "z_rot")
filtered_metrics = metrics.filter_metrics(foreground_placement_info_definition_id)
rotation_df = pd.DataFrame(filtered_metrics["rotation"].to_list(), columns=columns)
Getting Started¶
To get started using a sample project to generate synthetic data and to explore your dataset using this package please follow the SynthDet Documentation
Dataset Evaluation¶
To use the pre-compiled pipelines that allow you to evaluate the quality of synthetic dataset, you can follow the Dataset Evaluation Documentation.