SyntheticDataGenerator

The generation loop: randomize, optionally simulate physics, render, write.

Import

from telekinesis.illusion.core.synthetic_data_generator import SyntheticDataGenerator

Constructor

generator = SyntheticDataGenerator(context, randomizer, writer)

generator.generate(
    num_images=5,
    simulate_physics=False,
    min_simulation_time_range=(0.5, 1.0),
    max_simulation_time_range=(2.0, 5.0),
    save_blender_scene=False,
    clean_up_scene=True,
    render_max_retries=2,
)

It takes the Context holding the scene, the Randomizer describing how it varies, and a writer that serializes the rendered frames and annotations.

Parameters

Parameter	Type	Description
num_images	int	Number of scenes to render.
simulate_physics	bool	Whether to settle the objects with a physics simulation before rendering each scene.
min_simulation_time_range	tuple[float, float]	Range, in simulated seconds, the minimum simulation time is sampled from.
max_simulation_time_range	tuple[float, float]	Range, in simulated seconds, the maximum simulation time is sampled from.
check_object_interval	float	Interval, in simulated seconds, at which objects are checked for having come to rest.
object_stopped_location_threshold	float	Maximum location change, in meters, allowed between checks for an object to count as at rest.
object_stopped_rotation_threshold	float	Maximum rotation change, in radians, allowed between checks for an object to count as at rest.
substeps_per_frame	int	Physics substeps computed per simulation frame.
solver_iters	int	Solver iterations used by the physics simulation.
use_volume_com	bool	Compute the center of mass from object volume instead of mesh vertices.
verbose	bool	Verbose logging from the physics simulation.
save_blender_scene	bool	Save the first scene as a .blend file, for debugging.
clean_up_scene	bool	Clean up the scene after generation. Set to False when generating in a loop, as workers do.
render_max_retries	int	Retries after a transient render failure before the error is raised.

Rendering targets an NVIDIA GPU through OptiX, and segmentation output is enabled for category_id, instance, and name, which is what produces the instance masks in the annotations.

Next Steps

Workers

Drive a complete, sharded dataset run from a spec YAML instead of assembling the tree in Python.

Explore →

Merging and Converting

Merge shards into one annotation file and split it into train, valid, and test in COCO or YOLO format.

Explore →