Skip to content

README.md

Latest commit

 

History

History
100 lines (72 loc) · 3.39 KB

README.md

File metadata and controls

100 lines (72 loc) · 3.39 KB
Raw

Dataset Pipelines

This directory contains the processing pipelines for all datasets used in training and evaluation. Each subdirectory has its own detailed README.

Overview

Dataset Stages Source Output Conda envs
Grand Tour 2 Hugging Face Hub WebDataset (LiDAR + RGB + LoftUp) vernata
TartanGround 3 TartanAir server WebDataset (LiDAR + RGB + LoftUp) vernata
Waymo 2 Waymo Open Dataset (local TFRecords) WebDataset (LiDAR + camera + LoftUp) waymo → vernata

Grand Tour

→ Full README

Legged-robot dataset collected with a Hesai + Livox LiDAR stack and three HDR fisheye cameras. Raw sensor data is hosted on Hugging Face Hub as zarr archives, downloaded and aligned per mission, then processed into WebDataset shards with surface normals and LoftUp image features.

[Hugging Face Hub]
      │  align.py  (conda: vernata)
      ▼
[Local: per-mission .tar archives]
      │  to_webdataset_loftup.py  (conda: vernata)
      ▼
[Local: WebDataset shards]

Scripts: grandtour/align.py, grandtour/to_webdataset_loftup.py

TartanGround

→ Full README

Large-scale synthetic ground-robot dataset from the TartanAir simulator, covering 49 outdoor and indoor environments. Raw zip files are downloaded from the TartanAir server, processed into a Faiss-indexed intermediate format, then converted to WebDataset shards (with or without LoftUp features). Three dataset variants are built: see the full README.

[TartanAir server]
      │  download.py  (conda: vernata)
      ▼
[Local: raw zips]
      │  prepare_dataset.py  (conda: vernata)
      ▼
[Local: processed PCD + Faiss index]
      │  to_webdataset_loftup.py  (conda: vernata)
      ▼
[Local: WebDataset shards]

Scripts: tartanground/download.py, tartanground/prepare_dataset.py, tartanground/to_webdataset_loftup.py

Split configs: tartanground/splits.yaml (full, 43 train / 6 val), tartanground/splits_small.yaml (9 train / 6 val)

Prerequisite: The tartanairpy git submodule must be initialised before running Stages 1–2. See the TartanGround README for details.

Waymo

→ Full README

Autonomous-driving dataset from the Waymo Open Dataset. TFRecord segments are read from local disk, converted to an intermediate NumPy format, then processed into WebDataset shards with surface normals and LoftUp image features.

[Local: .tfrecord files]
      │  to_intermediate.py  (conda: waymo)
      ▼
[Local: per-segment NumPy archives]
      │  to_webdataset_loftup.py  (conda: vernata)
      ▼
[Local: WebDataset shards]

Scripts: waymo/to_intermediate.py, waymo/to_webdataset_loftup.py

Conda environments

Env Python Key packages Used for
vernata 3.12 torch+cu124, faiss-gpu-cu12, tartanairpy, huggingface-hub, webdataset, zarr Training, inference, LoftUp, all dataset processing except Waymo TFRecord parsing
waymo 3.10 tensorflow 2.11, waymo-open-dataset-tf-2-11-0, open3d Waymo TFRecord parsing only

→ Install vernata · → Install waymo