Skip to content

Datasets

Supported datasets

Class Layout Modalities Notes
SemanticKittiDataset synchronous lidar, labels Labels masked to lower 16 bits (strips instance IDs)
Rellis3DDataset synchronous lidar, labels Fixed Rellis-3D/ prefix in directory tree
Goose3DDataset synchronous lidar, labels Split directory appears at two levels
RawDataset asynchronous any channel Profile-free; channels & loaders from .apairo/channels.yaml. Loads apairo-extractor output
TartanKittiDataset asynchronous any TartanDrive v2 channel Fixed channel profile; auto-discovers channels via .apairo

Synchronous datasets

All three synchronous datasets share the same interface because they all extend ProfiledDataset.

import apairo

# SemanticKITTI
ds = apairo.SemanticKittiDataset("/data/semantic_kitti", keys=["lidar", "labels"])

# GOOSE
ds = apairo.Goose3DDataset("/data/goose", keys=["lidar", "labels"], split="train")

# Rellis-3D
ds = apairo.Rellis3DDataset("/data/rellis", keys=["lidar"])

Constructor parameters

Parameter Type Default Description
root_dir str \| Path -- Dataset root directory
keys list[str] \| None all non-optional keys Modalities to load
split str \| None None Filter by split directory ("train", "val", "test")

Expected directory layouts

<root>/
  sequences/
    00/
      velodyne/   000000.bin  000001.bin  ...
      labels/     000000.label  000001.label  ...
    01/
      ...

<root>/
  lidar/
    train/
      seq_001/   000000.bin  000001.bin  ...
      seq_002/   ...
    val/
      ...
  labels/
    train/
      seq_001/   000000.label  ...
    val/
      ...

<root>/
  Rellis-3D/
    00000/
      os1_cloud_node_kitti_bin/    000000.bin  ...
      os1_cloud_node_kitti_label/  000000.label  ...
    00001/
      ...

How synchronous datasets work

Each class declares a YAML structural profile that describes the directory layout, file extensions, dtypes, and any type transformations needed. apairo reads the profile at import time and handles discovery, loading, and casting automatically.

See YAML Profiles for the full specification, and Adding a Dataset to create your own.