nerfbaselines

class nerfbaselines.Cameras(*args, **kwargs)

Bases: GenericCameras[ndarray], Protocol

class nerfbaselines.Dataset

Bases: _IncompleteDataset

cameras: Cameras

image_paths: List[str]

image_paths_root: str

images: ndarray | List[ndarray]

images_points2D_xy: List[ndarray] | None

images_points3D_indices: List[ndarray] | None

mask_paths: List[str] | None

mask_paths_root: str | None

masks: ndarray | List[ndarray] | None

metadata: Dict

points3D_error: ndarray | None

points3D_rgb: ndarray | None

points3D_xyz: ndarray | None

class nerfbaselines.DatasetLoaderSpec

Bases: TypedDict

id: Required[str]

load_dataset_function: Required[str]

exception nerfbaselines.DatasetNotFoundError

Bases: Exception

class nerfbaselines.DatasetSpec

Bases: TypedDict

download_dataset_function: Required[str]

evaluation_protocol: str | EvaluationProtocolSpec

id: Required[str]

metadata: DatasetSpecMetadata

class nerfbaselines.DatasetSpecMetadata

Bases: TypedDict

default_metric: str

description: str

id: str

licenses: List[LicenseSpec]

link: str

metrics: List[str]

name: str

paper_authors: List[str]

paper_link: str

paper_title: str

scenes: List[Dict[str, str]]

class nerfbaselines.DownloadDatasetFunction(*args, **kwargs)

Bases: Protocol

class nerfbaselines.EvaluationProtocol(*args, **kwargs)

Bases: Protocol

accumulate_metrics(metrics: Iterable[Dict[str, float | int]]) → Dict[str, float | int]

evaluate(predictions: Dict[str, ndarray], dataset: Dataset) → Dict[str, float | int]

get_name() → str

render(method: Method, dataset: Dataset, *, options: RenderOptions | None = None) → Dict[str, ndarray]

class nerfbaselines.EvaluationProtocolSpec

Bases: TypedDict

evaluation_protocol_class: Required[str]

id: Required[str]

class nerfbaselines.GenericCameras(*args, **kwargs)

Bases: Protocol[TTensor_co]

apply(fn: Callable[[TTensor_co, str], TTensor]) → GenericCameras[TTensor]

property camera_models: TTensor_co

Camera models, [N]

classmethod cat(values: Sequence[Self]) → Self

property distortion_parameters: TTensor_co

Distortion parameters, [N, num_params]

property image_sizes: TTensor_co

Image sizes, [N, 2]

property intrinsics: TTensor_co

Intrinsics, [N, (fx,fy,cx,cy)]

item() → Self

Returns a single camera if there is only one. Otherwise raises an error.

property metadata: TTensor_co | None

Metadata, [N, …]

property nears_fars: TTensor_co | None

Near and far planes, [N, 2]

property poses: TTensor_co

Camera-to-world matrices, [N, (R, t)]

replace(**changes) → Self

class nerfbaselines.LicenseSpec

Bases: TypedDict

name: Required[str]

url: str

class nerfbaselines.LoadDatasetFunction(*args, **kwargs)

Bases: Protocol

class nerfbaselines.Logger(*args, **kwargs)

Bases: Protocol

add_embedding(tag: str, embeddings: ndarray, step: int, *, images: List[ndarray] | None = None, labels: None | List[Dict[str, str]] | List[str] = None) → None

add_event(step: int) → ContextManager[LoggerEvent, bool | None]

add_image(tag: str, image: ndarray, step: int, *, display_name: str | None = None, description: str | None = None) → None

add_scalar(tag: str, value: float | int, step: int) → None

add_text(tag: str, text: str, step: int) → None

class nerfbaselines.LoggerEvent(*args, **kwargs)

Bases: Protocol

add_embedding(tag: str, embeddings: ndarray, *, images: List[ndarray] | None = None, labels: None | List[Dict[str, str]] | List[str] = None) → None

add_histogram(tag: str, values: ndarray, *, num_bins: int | None = None) → None

add_image(tag: str, image: ndarray, display_name: str | None = None, description: str | None = None, **kwargs) → None

add_plot(tag: str, *data: ndarray, axes_labels: Sequence[str] | None = None, title: str | None = None, **kwargs) → None

add_scalar(tag: str, value: float | int) → None

add_text(tag: str, text: str) → None

class nerfbaselines.LoggerSpec

Bases: TypedDict

id: Required[str]

logger_class: Required[str]

class nerfbaselines.Method(*, checkpoint: str | None = None, train_dataset: Dataset | None = None, config_overrides: Dict[str, Any] | None = None)

Bases: Protocol

abstract get_info() → ModelInfo

Get method defaults for the trainer.

Returns:

Method info.

abstract classmethod get_method_info() → MethodInfo

Get method info needed to initialize the datasets.

Returns:

Method info.

get_train_embedding(index: int) → ndarray | None

Get the embedding for the given image index.

Parameters:

index – Image index.

Returns:

Image embedding.

optimize_embedding(dataset: Dataset, *, embedding: ndarray | None = None) → OptimizeEmbeddingOutput

Optimize embedding for a single image (passed as a dataset with a single image).

Parameters:

• dataset – A dataset with a single image.

• embeddings – Optional initial embedding.

abstract render(camera: Cameras, *, options: RenderOptions | None = None) → Dict[str, ndarray]

Render single image.

Parameters:

• camera – Camera from which the scene is to be rendered.

• options – Optional rendering options.

abstract save(path: str) → None

Save model.

Parameters:

path – Path to save.

abstract train_iteration(step: int) → Dict[str, float]

Train one iteration.

Parameters:

step – Current step.

class nerfbaselines.MethodInfo

Bases: TypedDict

can_resume_training: bool

method_id: Required[str]

required_features: FrozenSet[Literal['color', 'points3D_xyz', 'points3D_rgb', 'images_points3D_indices', 'images_points2D_xy', 'points3D_error']]

supported_camera_models: FrozenSet[Literal['pinhole', 'opencv', 'opencv_fisheye', 'full_opencv']]

supported_outputs: Tuple[str | RenderOutputType, ...]

viewer_default_resolution: int | Tuple[int, int]

class nerfbaselines.MethodSpec

Bases: TypedDict

apptainer: NotRequired[Any]

backends_order: List[Literal['conda', 'docker', 'apptainer', 'python']]

conda: NotRequired[Any]

docker: NotRequired[Any]

id: Required[str]

implementation_status: Dict[str, Literal['working', 'reproducing', 'not-working', 'working-not-reproducing']]

metadata: Dict[str, Any]

method_class: Required[str]

output_artifacts: Dict[str, OutputArtifact]

presets: Dict[str, Dict[str, Any]]

required_features: List[Literal['color', 'points3D_xyz', 'points3D_rgb', 'images_points3D_indices', 'images_points2D_xy', 'points3D_error']]

supported_camera_models: List[Literal['pinhole', 'opencv', 'opencv_fisheye', 'full_opencv']]

supported_outputs: List[str | RenderOutputType]

class nerfbaselines.ModelInfo

Bases: TypedDict

batch_size: int

can_resume_training: bool

eval_batch_size: int

hparams: Dict[str, Any]

loaded_checkpoint: str | None

loaded_step: int | None

method_id: Required[str]

num_iterations: Required[int]

required_features: FrozenSet[Literal['color', 'points3D_xyz', 'points3D_rgb', 'images_points3D_indices', 'images_points2D_xy', 'points3D_error']]

supported_camera_models: FrozenSet

supported_outputs: Tuple[str | RenderOutputType, ...]

viewer_default_resolution: int | Tuple[int, int]

class nerfbaselines.OptimizeEmbeddingOutput

Bases: TypedDict

embedding: Required[ndarray]

metrics: NotRequired[Dict[str, Sequence[float]]]

render_output: NotRequired[Dict[str, ndarray]]

class nerfbaselines.OutputArtifact

Bases: TypedDict

link: Required[str]

class nerfbaselines.RenderOptions

Bases: TypedDict

embedding: ndarray | None

output_type_dtypes: Dict[str, str]

outputs: Tuple[str, ...]

class nerfbaselines.RenderOutputType

Bases: TypedDict

name: Required[str]

type: Literal['color', 'depth']

class nerfbaselines.Trajectory

Bases: TypedDict

appearances: NotRequired[List[TrajectoryFrameAppearance]]

camera_model: Literal['pinhole', 'opencv', 'opencv_fisheye', 'full_opencv']

fps: float

frames: List[TrajectoryFrame]

image_size: Tuple[int, int]

source: Any

class nerfbaselines.TrajectoryFrame

Bases: TypedDict

appearance_weights: NotRequired[ndarray]

intrinsics: ndarray

pose: ndarray

class nerfbaselines.TrajectoryFrameAppearance

Bases: TypedDict

embedding: ndarray | None

embedding_train_index: int | None

class nerfbaselines.UnloadedDataset

Bases: _IncompleteDataset

cameras: Cameras

image_paths: List[str]

image_paths_root: str

images: NotRequired[ndarray | List[ndarray] | None]

images_points2D_xy: List[ndarray] | None

images_points3D_indices: List[ndarray] | None

mask_paths: List[str] | None

mask_paths_root: str | None

masks: ndarray | List[ndarray] | None

metadata: Dict

points3D_error: ndarray | None

points3D_rgb: ndarray | None

points3D_xyz: ndarray | None

nerfbaselines.build_method_class(spec: MethodSpec, backend: Literal['conda', 'docker', 'apptainer', 'python'] | None = None)

Build a method class from a method spec. It automatically selects the backend based on the method spec if none is provided.

Parameters:

• spec – Method spec

• backend – Backend name

nerfbaselines.camera_model_from_int(i: int) → Literal['pinhole', 'opencv', 'opencv_fisheye', 'full_opencv']

nerfbaselines.camera_model_to_int(camera_model: Literal['pinhole', 'opencv', 'opencv_fisheye', 'full_opencv']) → int

nerfbaselines.get_dataset_loader_spec(id: str) → DatasetLoaderSpec

Get a dataset loader specification by registered dataset loader ID.

Parameters:

id – Dataset loader ID

Returns:

Dataset loader specification

nerfbaselines.get_dataset_spec(id: str) → DatasetSpec

Get a dataset specification by registered dataset ID.

Parameters:

id – Dataset ID

Returns:

Dataset specification

nerfbaselines.get_evaluation_protocol_spec(id: str) → EvaluationProtocolSpec

Get an evaluation protocol specification by registered evaluation protocol ID.

Parameters:

id – Evaluation protocol ID

Returns:

Evaluation protocol specification

nerfbaselines.get_logger_spec(id: str) → LoggerSpec

Get a logger specification by registered logger ID.

Parameters:

id – Logger ID

Returns:

Logger specification

nerfbaselines.get_method_spec(id: str) → MethodSpec

Get a method by method ID.

Parameters:

id – Method ID

Returns:

Method spec

nerfbaselines.get_supported_dataset_loaders() → FrozenSet[str]

Get all supported dataset loaders. The loaders are sorted by priority.

Returns:

List of dataset loader IDs (sorted by priority)

nerfbaselines.get_supported_datasets() → FrozenSet[str]

Get all supported datasets.

Returns:

Set of dataset IDs

nerfbaselines.get_supported_evaluation_protocols() → FrozenSet[str]

Get all supported evaluation protocols.

Returns:

Set of evaluation protocol IDs

nerfbaselines.get_supported_loggers() → FrozenSet[str]

Get all supported loggers.

Returns:

Set of logger IDs

nerfbaselines.get_supported_methods(backend_name: Literal['conda', 'docker', 'apptainer', 'python'] | None = None) → FrozenSet[str]

Get all supported methods. Optionally, filter the methods that support a specific backend.

Parameters:

backend_name – Backend name

Returns:

Set of method IDs

nerfbaselines.load_checkpoint(checkpoint: str, *, backend: Literal['conda', 'docker', 'apptainer', 'python'] | None = None) → Generator[Tuple[Method, Dict], None, None]

This is a utility function to open the checkpoint directory, mount it, start the backend, build the model class and load the checkpoint. The checkpoint can be a local path, a remote path or a path inside a zip file. The function returns a context manager that yields the model instance and nb-info.

Parameters:

• checkpoint – Path to the checkpoint. Can be a local path or a remote path. Can also be a path inside a zip file.

• backend – Backend name

Returns:

Context manager that yields a tuple of model instance and the nb-info dictionary.

nerfbaselines.new_cameras(*, poses: ndarray, intrinsics: ndarray, camera_models: ndarray, image_sizes: ndarray, distortion_parameters: ndarray | None = None, nears_fars: ndarray | None = None, metadata: ndarray | None = None) → Cameras

nerfbaselines.new_dataset(*, cameras: Cameras, image_paths: Sequence[str], image_paths_root: str | None = None, images: ndarray | List[ndarray] | None = None, mask_paths: Sequence[str] | None = None, mask_paths_root: str | None = None, masks: ndarray | List[ndarray] | None = None, points3D_xyz: ndarray | None = None, points3D_rgb: ndarray | None = None, points3D_error: ndarray | None = None, images_points3D_indices: Sequence[ndarray] | None = None, images_points2D_xy: Sequence[ndarray] | None = None, metadata: Dict | None = None) → UnloadedDataset | Dataset

nerfbaselines.register(spec: MethodSpec | DatasetSpec | DatasetLoaderSpec | EvaluationProtocolSpec | LoggerSpec) → None

Register a method, dataset, logger, or evaluation protocol spec.

Parameters:

spec – Spec to register (MethodSpec, DatasetSpec, DatasetLoaderSpec, EvaluationProtocolSpec, LoggerSpec)