Synthetic Homes

Unity Synthetic Homes is a dataset generator and accompanying large-scale dataset of photorealistic randomized home interiors, built for training computer vision models such as object detection, scene understanding, and monocular depth estimation.

The application performs a wide variety of randomizations to maximize the diversity of generated datasets. These include materials, furniture type and configuration, sunlight angle and temperature, day/night switching, interior lighting temperature, camera angles, clutter, skybox, door and curtain animations, and more. By providing a configuration file, you control customize many of these elements, enabling you to tune them to your liking.

This applications was made using the Unity Perception package, which provides tools for generating randomized synthetic CV datasets with a wide variety of ground-truth annotations.

Interior lighting in homes is complex and is difficult to replicate with traditional raster-based methods. We used Unity’s multi-bounce path tracing to accomplish physically accurate global illumination and reflections. This accuracy can help bridge the so called “Sim2Real gap”, improving a model’s ability to perform well in the real world after training on synthetic data.

Included Label Types

• Instance and semantic segmentation
• 2D bounding boxes
• Occlusion percentage of each object
• Depth
• Normals
• Pixel word position
• Camera position and properties

How to use the 100k dataset

We built a dataset of 100k home interior images and annotations in SOLO format. It is hosted at https://console.cloud.google.com/storage/browser/unity-cv-dataset-customer-examples/100k-dataset .

A 500 image sample dataset can be found here.

To download the entire 100k datset (~1.1TB), follow these instructions:

1. Install gsutil
2. Download the dataset via the command line

gsutil -m cp -r gs://unity-cv-dataset-customer-examples/100k-dataset/ ./

Analyzing and Visualizing Datasets

Datasets are generated in the SOLO format. For information on how to explore and analyze SOLO datasets, check out the endpoint and schema documentation pages on the Perception repository.

Dataset Generator

Head over to the Releases page and download the latest build. Once the archive is extracted, you can simply double click SyntheticHomes.exe to run the application with its default settings. A window will be opened and frames will start to be randomized and rendered. Each final frame will take a while to render as we accumulate multiple frames to achieve high quality path traced results.

By default, the generated dataset will be located at %USERPROFILE%\AppData\LocalLow\UnityTechnologies\SyntheticHomes. You can browse through the dataset while the generator is running.

Alternatively, you can supply command line arguments and an optional JSON configuration file to modify various settings of your run.

System Requirements

• Windows 10 or 11 64-bit
• DX12 with DXR compatible GPU for path tracing (list available here)
  • The application will fall back to rasterization if DXR capable hardware is not detected.

Command Line Arguments

• --scenario-config-file=<path-to-json-file>

  • Supply a configuration file to control various parameters related to rendering and randomization. See below for Instructions.

• --resolution=<width>x<height>

  • Set the resolution of the output images. Example usage: --resolution=1600x1200

• --output-path <path>

  • Specify output folder

Configuration File

Several aspects of the dataset generation can be controlled using a JSON config file that is provided to the application. A sample config file is provided here.

The configuration file includes three main sections: (a) constants, (b) sensors, and (c)randomizers.

In the constants block, you specify:

• iterationCount: The number of Iterations the simulation should run for. Each Iteration produces one annotated frame. E.g. to generate 1000 images in your dataset, you should set thus value to 1000.
• randomSeed: The seed used for randomization. This generator uses Perception's randomization framework, which generates deterministic random numbers based on a provided seed. This helps you replicate datasets by keeping your seed value and randomization settings unchanged.

In the sensors block, you can enable or disable sensors and their Labelers. This project contains a single sensor of type PerceptionCamera, so there is no point in disabling it. However, you can disable Labelers that you do not need to reduce the size of the output dataset by modifying the enabled field of each Labeler in the labelers array of the PerceptionCamera. If no JSON config is provided, all Labelers are activated.

In the randomizers block of the JSON config file, you control the behavior of Perception Randomizers included in the project. Each Randomizer has one JSON block in this file. randomizerId denotes the name of the Randomizer, and items are the list of settings that can be changed. To change the behavior for each item, what you will need to modify is the value blok nested inside. In addition, some Randomizers can be completely disabled. We will now go through all the available Randomizers and their settings:

• ScenarioSettings: Controls high level settings for the simulation.
  • includeSingleFamilyVariantA, includeSingleFamilyVariantB, includeSingleFamilyVariantC, includeMultiFamilyVariantA: The generator includes four house shells. The bool value of true or false for each of these specifies whether they should be used by the generator. The first three listed here are single story houses, while the fourth is a multi-story one.
  • includeBathrooms, includeKitchens, includeBedrooms, includeDiningRooms, includeLivingRooms, includeHallsAndStairs, includeLaundryRooms: Set the bool value of these to true or false to specify whether a certain type of room should be included or not.
  • usePathTracing: Whether hardware accelerated path tracing should be enabled. Note that this only works with DXR compatible GPUs.
  • pathTracingSamples: Sets the number of path traced frames to accumulate for the final image. There is a progress bar at the bottom of the screen view which indicates the current accumulation with respect to this value. Higher values lead to sharper and higher quality images.
  • pathTracingMaxDepth: Sets the maximum number of light bounces in each path.
  • denoiserType: We use a denoising algorithm to achieve the final image after all the path tracing samples have been accumulated. In our experience, even with very large values for pathTracingSamples (above 1000), denoising is still needed for noise free output. By default, we use the Intel Open Image algorithm which seems to yield generally better results. We recommend you experiment with various values for number of samples and denoiser type to achieve your desired outcomes.
    • Acceptable values for the str field for denoiserType:
      • intel: Use the Intel Open Image denoiser
      • nvda: Use the NVIDA Optix denoiser
      • none: Disable denoising
  • useAovsForDenoising: When either nvda or intel are selected above, this flag specifies whether the texture and normal information on the materials is used to enhance the output of the denoising algorithm. This can substantially improve the look of most surfaces, but can have a negative detail crushing effect on surfaces that are visible behind other transparent objects, such as bathtub walls behind shower doors.
  • enableTemporalDenoising: Controls whether temporal information is used in denoising. Only applies to the nvda option above.
  • enableBloom: Enables a bloom effect on very bright highlights. Note that enabling this can make it look like the camera's view is completely black until the full frame is accumulated.
• SceneRandomizer: Each of the house shells mentioned earlier is in a separate Unity Scene. This Randomizer handles the switching of Scenes and thus house shells.
  • iterationsPerScene: Set the num field of this block to the number of Iterations you would like each house shell to be used in. Your iterationCount set in the constants block above divided by this value equals the number of house shells your simulation has time to visit before it ends.
• VantagePointsCameraPlacementRandomizer: The camera is randomized by selecting one of several available vantage points.
  • pitch: The angle (in degrees) at which the camera is looking up or down from the horizon. Values are uniformly sampled from the given range. Positive values result in the camera looking up and negative values make the camera look down. Change the min and max values nested inside this block to modify camera pitch. The valid range is [-180, 180].
  • yaw: The horizontal angle (in degrees) between the camera's look direction and the line between the camera's position and the center of the room. Values are uniformly sampled from the given range. Positive values result in the camera looking to the right and negative values make it look to the left. Change the min and max values nested inside this block to modify camera yaw. The valid range is [-180, 180].
  • roll: Degrees of rotation around the forward axis of the camera. Non-zero values lead to non-straight horizon lines. Values are uniformly sampled from the given range. Change the min and max values nested inside this block to modify camera roll. The valid range is [-180, 180].
  • heightFromFloor: Height of the camera from the floor, in meters. Sampled uniformly from the min to max range specified here. The ceiling is typically at about 2.7 meters or higher. When placing the camera near the floor, make sure you provide a range of pitch angles that cause it to look up, otherwise it can look through the floor. Similarly, if your camera is very high, make it look down.
  • fovRange: Range of angles to use for the camera's field of view. Sampled uniformly from the min to max range specified here.
• SplitGrammarRandomizer: Generates and randomizes the type and arrangement of furniture. Set the enabled value to true or false to enable or disable furniture completely.
• SkyBoxRandomizer: Randomizes the HDRI skybox. If disabled, a single skybox will be used in all images.
• CameraPostProcessingRandomizer: Controls a number of post processing effects on the camera.
  • blurOnProbability: The probability of lens blur being enabled in a frame. Values are uniformly sampled from the given min to max range. The valid range is [0, 1].
  • blurIntensityParameter: Range of of intensity of the blur effect, when it is randomly turned on based on the probability above. Values are uniformly sampled from the given min to max range. The valid range is [0, 1].
  • contrastParameter and saturationParameter: Values are uniformly sampled from the given min to max range. The valid range is [-100, 100]. Note that numbers close to 100 or -100 can lead to completely unusable images.
• DoorOpeningRandomizer: Randomizes the degree to which doors are open. If disabled, all doors will be in a fully closed state.
• MaterialSwapperRandomizer: Randomizes the materials on a large portion of the surfaces. If disabled, no materials will be randomized and walls will have a typical white color.
• ClutterSwapperRandomizer: Generates and randomizes the occasional clutter on bathroom floors. If disabled, no clutter will be generated.
• CustomSunAngleRandomizer: Randomizes the direction and temperature of sunlight based on the given parameters.
  • warmLimit and coldLimit: The warmest (for sundown and sunrise) and coldest (for mid-day) color temperatures to use.
  • dayTimeBegin and dayTimeEnd: Specifies what hours of the day (e.g. 8 and 17) correspond to the above warmLimit value. The coldLimit value will be mapped to the hour exactly in-between these two values (in our example 12:30 pm) Valid range for both is [0, 24] and dayTimeBegin should be smaller than dayTimeEnd.
  • hourRange1 and hourRange2: You have the option of providing two ranges of hours. When randomizing the sun, one of these ranges will be picked, then a random hour in the picked range will be selected, and then the sun's direction and temperature will be set based on the chosen hour. Valid range for both is [0, 24] and both ranges should fall within the range of daylight hours specified above using dayTimeBegin and dayTimeEnd.
• DayNightSwitcherRandomizer: Decides whether day or night time illumination is used in each frame, and can also randomize camera exposure. At daytime, the sunlight and the ambient light coming in from the windows are the only sources of illumination. At night time, the sun is deactivated and interior light fixtures are turned on.
  • dayProbability: The probability that day time is selected. A value of 1 for the num field here results in all images being day time.
  • onlySelectDaytimeIfRoomHasWindows: Some rooms do not have any windows and would thus be completely dark in day time lighting. If this flag is enabled, day time would never be selected when the camera is in those rooms.
  • randomExposure: Enable this to introduce some randomization into the overall exposure of the camera, resulting in darker or brighter images.
  • exposureRange: The range of exposures to use when the above randomExposure flag is enabled. The valid range is [0, 1]. Smaller numbers generate darker images.
• LightFixtureController: Randomizes the temperature of interior light fixtures, which is uniformly sampled from the provided min and max values. Each light fixture is randomized individually, resulting in a variety of light temperatures in the house.

License

• License

Citation

If you find SynthHomes useful, consider citing it using:

@misc{SyntHomes,
    title={Unity SynthHomes: A Synthetic Home Interior Dataset Generator},
    author={{Unity Technologies}},
    howpublished={\url{https://github.com/Unity-Technologies/SynthHomes}},
    year={2022}
}