vision3d.datasets

3D object detection dataset loaders.

Functions

collate_fn(batch)

Collate a batch of (inputs, targets) without stacking.

Classes

CameraInputs	Camera-only sample: images, intrinsics, and extrinsics always present.
FusionInputs	Fusion sample: lidar plus multi-camera, every field present.
Kitti3D(root[, train, transforms, download])	KITTI 3D Dataset.
LidarInputs	Lidar-only sample: points always present.
NuScenes3D(root[, version, split, transforms])	nuScenes 3D object detection dataset.
SampleInputs	Per-frame model inputs; base type with all fields optional.
SampleTargets	Per-frame ground-truth annotations.

class vision3d.datasets.CameraInputs

Bases: SampleInputs

Camera-only sample: images, intrinsics, and extrinsics always present.

class vision3d.datasets.FusionInputs

Bases: SampleInputs

Fusion sample: lidar plus multi-camera, every field present.

class vision3d.datasets.Kitti3D(root, train=True, transforms=None, download=False)

Bases: Dataset[tuple[FusionInputs, SampleTargets | None]]

KITTI 3D Dataset.

Returns samples in lidar frame (X-forward, Y-left, Z-up), converting from KITTI’s camera-frame annotations automatically.

Parameters:

• root (str or pathlib.Path) –

  Root directory where data is downloaded to. Expects the following folder structure if download=False:

  <root>
      └── Kitti3D/
          └── raw/
              ├── training/
              |   ├── velodyne/
              |   ├── label_2/
              |   ├── calib/
              |   └── image_2/
              └── testing/
                  ├── velodyne/
                  ├── calib/
                  └── image_2/
  

• train (bool, optional) – Use train split if true, else test split. Defaults to True.

• transforms (Callable, optional) – A function/transform that takes input sample and its target as entry and returns a transformed version.

• download (bool, optional) – If true, downloads the dataset from the internet and puts it in root directory. If dataset is already downloaded, it is not downloaded again.

download()

Download the KITTI data if it doesn’t exist already.

Return type:

None

class vision3d.datasets.LidarInputs

Bases: SampleInputs

Lidar-only sample: points always present.

class vision3d.datasets.NuScenes3D(root, version='v1.0-mini', split='train', transforms=None)

Bases: Dataset[tuple[FusionInputs, SampleTargets]]

nuScenes 3D object detection dataset.

Returns samples in the global frame with annotations as BoundingBoxes3D in XYZLWHY format (yaw extracted from quaternion). Multi-camera images, intrinsics, and extrinsics are returned for all 6 cameras.

Requires the nuscenes-devkit package.

Parameters:

• root (str or pathlib.Path) – Root directory of the nuScenes dataset.

• version (str) – Dataset version. Default: "v1.0-mini".

• split (str) – One of "train" or "val". Default: "train".

• transforms (Callable, optional) – A function/transform that takes input sample and its target as entry and returns a transformed version.

class vision3d.datasets.SampleInputs

Bases: TypedDict

Per-frame model inputs; base type with all fields optional.

Fields are ReadOnly so dataset-specific subclasses can narrow them from NotRequired to Required.

Variables:

• points (vision3d.tensors._point_cloud_3d.PointCloud3D) – Lidar point cloud for the frame.

• images (vision3d.tensors._camera.CameraImages) – Multi-camera image tensor, one row per camera.

• extrinsics (vision3d.tensors._camera.CameraExtrinsics) – Lidar-to-camera transforms, one row per camera.

• intrinsics (vision3d.tensors._camera.CameraIntrinsics) – Per-camera pinhole intrinsic matrices.

class vision3d.datasets.SampleTargets

Bases: TypedDict

Per-frame ground-truth annotations.

Variables:

• boxes (vision3d.tensors._bounding_boxes_3d.BoundingBoxes3D) – 3D bounding boxes in the lidar frame.

• labels (torch.Tensor) – Integer class labels, one per box.

vision3d.datasets.collate_fn(batch)

Collate a batch of (inputs, targets) without stacking.

Variable-size tensors (point clouds, bounding boxes) cannot be stacked into a single tensor. This collate function groups them as tuples, matching torchvision’s detection collate pattern.

Parameters:

batch (list[tuple[Any, Any]]) – List of (inputs, targets) from the dataset.

Returns:

Tuple of (inputs_tuple, targets_tuple).

Return type:

tuple[tuple[Any, …], tuple[Any, …]]