l5kit.rasterization package

class l5kit.rasterization.BoxRasterizer(render_context: l5kit.rasterization.render_context.RenderContext, filter_agents_threshold: float, history_num_frames: int, render_ego_history: bool = True)

Bases: l5kit.rasterization.rasterizer.Rasterizer

num_channels() → int
rasterize(history_frames: numpy.ndarray, history_agents: List[numpy.ndarray], history_tl_faces: List[numpy.ndarray], agent: Optional[numpy.ndarray] = None) → numpy.ndarray

Generate raster image by rendering Ego & Agents as bounding boxes on the raster image. Ego & Agents from different past frame are rendered at different image channel.

Parameters

  • history_frames – A list of past frames to be rasterized

  • history_agents – A list of agents from past frames to be rasterized

  • history_tl_faces – A list of traffic light faces from past frames to be rasterized

  • agent – The selected agent to be rendered as Ego, if it is None the AV will be rendered as Ego

Returns

An raster image of size [2xN] with Ego & Agents rendered as bounding boxes, where N is number of history frames

to_rgb(in_im: numpy.ndarray, **kwargs: dict) → numpy.ndarray

This function is used to get an rgb image where agents further in the past have faded colors.

Parameters

  • in_im – The output of the rasterize function

  • kwargs – This can be used for additional customization (such as colors)

Returns

An RGB image with agents and ego coloured with fading colors

class l5kit.rasterization.Rasterizer

Bases: abc.ABC

Base class for something that takes a single state of the world, and outputs a (multi-channel) image.

abstract num_channels() → int
abstract rasterize(history_frames: numpy.ndarray, history_agents: List[numpy.ndarray], history_tl_faces: List[numpy.ndarray], agent: Optional[numpy.ndarray] = None) → numpy.ndarray
abstract to_rgb(in_im: numpy.ndarray, **kwargs: dict) → numpy.ndarray
class l5kit.rasterization.RenderContext(raster_size_px: numpy.ndarray, pixel_size_m: numpy.ndarray, center_in_raster_ratio: numpy.ndarray, set_origin_to_bottom: bool)

Bases: object

raster_from_world(position_m: numpy.ndarray, angle_rad: float) → numpy.ndarray

Return a matrix to convert a pose in world coordinates into raster coordinates

Parameters

  • render_context (RenderContext) – the context for rasterisation

  • position_m (np.ndarray) – XY position in world coordinates

  • angle_rad (float) – rotation angle in world coordinates

Returns

a transformation matrix from world coordinates to raster coordinates

Return type

(np.ndarray)

class l5kit.rasterization.SatBoxRasterizer(render_context: l5kit.rasterization.render_context.RenderContext, filter_agents_threshold: float, history_num_frames: int, map_im: numpy.ndarray, world_to_aerial: numpy.ndarray, interpolation: int = 1, render_ego_history: bool = True)

Bases: l5kit.rasterization.rasterizer.Rasterizer

Combine a Satellite and a Box Rasterizers into a single class

num_channels() → int
rasterize(history_frames: numpy.ndarray, history_agents: List[numpy.ndarray], history_tl_faces: List[numpy.ndarray], agent: Optional[numpy.ndarray] = None) → numpy.ndarray
to_rgb(in_im: numpy.ndarray, **kwargs: dict) → numpy.ndarray
class l5kit.rasterization.SatelliteRasterizer(render_context: l5kit.rasterization.render_context.RenderContext, map_im: numpy.ndarray, world_to_aerial: numpy.ndarray, interpolation: int = 1)

Bases: l5kit.rasterization.rasterizer.Rasterizer

This rasterizer takes a satellite image in its constructor and a transform from world coordinates to this image. When you call rasterize, it will return a crop around the agent of interest with the agent’s forward vector pointing right for the current timestep.

num_channels() → int
rasterize(history_frames: numpy.ndarray, history_agents: List[numpy.ndarray], history_tl_faces: List[numpy.ndarray], agent: Optional[numpy.ndarray] = None) → numpy.ndarray
to_rgb(in_im: numpy.ndarray, **kwargs: dict) → numpy.ndarray
class l5kit.rasterization.SemBoxRasterizer(render_context: l5kit.rasterization.render_context.RenderContext, filter_agents_threshold: float, history_num_frames: int, semantic_map_path: str, world_to_ecef: numpy.ndarray, render_ego_history: bool = True)

Bases: l5kit.rasterization.rasterizer.Rasterizer

Combine a Semantic Map and a Box Rasterizers into a single class

num_channels() → int
rasterize(history_frames: numpy.ndarray, history_agents: List[numpy.ndarray], history_tl_faces: List[numpy.ndarray], agent: Optional[numpy.ndarray] = None) → numpy.ndarray
to_rgb(in_im: numpy.ndarray, **kwargs: dict) → numpy.ndarray
class l5kit.rasterization.SemanticRasterizer(render_context: l5kit.rasterization.render_context.RenderContext, semantic_map_path: str, world_to_ecef: numpy.ndarray)

Bases: l5kit.rasterization.rasterizer.Rasterizer

Rasteriser for the vectorised semantic map (generally loaded from json files).

num_channels() → int
rasterize(history_frames: numpy.ndarray, history_agents: List[numpy.ndarray], history_tl_faces: List[numpy.ndarray], agent: Optional[numpy.ndarray] = None) → numpy.ndarray
render_semantic_map(center_in_world: numpy.ndarray, raster_from_world: numpy.ndarray, tl_faces: numpy.ndarray) → numpy.ndarray

Renders the semantic map at given x,y coordinates.

Parameters

  • center_in_world (np.ndarray) – XY of the image center in world ref system

  • raster_from_world (np.ndarray) –

Returns

RGB raster

Return type

np.ndarray

to_rgb(in_im: numpy.ndarray, **kwargs: dict) → numpy.ndarray
class l5kit.rasterization.StubRasterizer(render_context: l5kit.rasterization.render_context.RenderContext)

Bases: l5kit.rasterization.rasterizer.Rasterizer

This rasterizer doesn’t actually do anything, it returns an all-black image. Useful for testing.

num_channels() → int
rasterize(history_frames: numpy.ndarray, history_agents: List[numpy.ndarray], history_tl_faces: List[numpy.ndarray], agent: Optional[numpy.ndarray] = None) → numpy.ndarray

Just return an empty raster_size x 3 black image

to_rgb(in_im: numpy.ndarray, **kwargs: dict) → numpy.ndarray

Return a completely black image.

l5kit.rasterization.build_rasterizer(cfg: dict, data_manager: l5kit.data.local_data_manager.DataManager) → l5kit.rasterization.rasterizer.Rasterizer

Factory function for rasterizers, reads the config, loads required data and initializes the correct rasterizer.

Parameters

  • cfg (dict) – Config.

  • data_manager (DataManager) – Datamanager that is used to require files to be present.

Raises

  • NotImplementedError – Thrown when the map_type read from the config doesn’t have an associated rasterizer

  • type in this factory function. If you have custom rasterizers, you can wrap this function in your own factory –

  • function and catch this error. –

Returns

Rasterizer initialized given the supplied config.

Return type

Rasterizer

l5kit.rasterization.get_sat_image_crop(sat_image: numpy.ndarray, crop_size: Union[Tuple[int, int], numpy.ndarray], sat_pixel_translation: numpy.ndarray, yaw: Optional[float] = None) → numpy.ndarray

Crops input satellite such that sat_pixel_translation is centered in the image.

Parameters

  • sat_image (np.ndarray) – satellite image

  • crop_size (Union[Tuple[int, int], np.ndarray]) – size of desired crop in pixels

  • sat_pixel_translation (np.ndarray) – 2D or 3D vector where to center the cropped image in pixels.

Keyword Arguments

yaw (Optional[float]) – yaw in radians, None or 0 means no rotation is applied to the output image. default: {None})

Returns

a crop of input sat_image

Return type

(np.ndarray)

l5kit.rasterization.get_sat_image_crop_scaled(sat_image: numpy.ndarray, crop_size: Union[Tuple[int, int], numpy.ndarray], sat_pixel_translation: numpy.ndarray, yaw: Optional[float] = None, sat_pixel_scale: float = 1.0, pixel_size: float = 1.0, interpolation: int = 1) → numpy.ndarray
Calls get_sat_image_crop (see that function’s docs for further details), and rescales taking

into account a desired pixel size.

Example

Desired crop_size is 200x200, and pixel_size is 0.5: we want an image that corresponds to 100x100 meters. This means it extracts a 33x33 image and scales it up to 200x200.

Parameters

  • sat_image (np.ndarray) – satellite image

  • crop_size (Union[Tuple[int, int], np.ndarray]) – size of desired crop in pixels

  • sat_pixel_translation (np.ndarray) – 2D or 3D vector where to center the cropped image in pixels.

Keyword Arguments

  • yaw (Optional[float]) – yaw in radians, 0 means no rotation is applied, which generally means up is North. default: {None})

  • sat_pixel_scale (float) – A sat_pixel_scale of 3.0 would means that every pixel in the sat

  • (default (image corresponds to 3m in the real world.) – {1.0})

  • pixel_size (float) – [description] (default: {1.0})

  • interpolation (int) – [description] (default: {cv2.INTER_LINEAR})

Returns

a crop of input sat_image

Return type

(np.ndarray)

l5kit.rasterization.get_sat_image_crop_scaled_from_ecef(sat_image: numpy.ndarray, crop_size: Union[Tuple[int, int], numpy.ndarray], ecef_translation: numpy.ndarray, ecef_to_sat: numpy.ndarray, **kwargs: Any) → numpy.ndarray

Utility function, calls get_sat_image_crop_scaled, see that function for more details on additional keyword arguments (such as yaw).

Parameters

  • sat_image (np.ndarray) – satellite image

  • crop_size (Union[Tuple[int, int], np.ndarray]) – size of desired crop in pixels

  • ecef_translation (np.ndarray) – 2D or 3D vector where to center the cropped image

  • ecef_to_sat (np.ndarray) – transform from ECEF to satellite image space

Returns

np.ndarray – a crop of satellite_image

Submodules