Datasets:

Tera-AI
/

STRIDE

license: cc-by-nc-4.0
task_categories:
  - image-to-text
dataset_info:
  features:
    - name: image
      dtype: image
    - name: file_name
      dtype: string
    - name: path_index
      dtype: int64
    - name: node_index
      dtype: int64
    - name: split
      dtype: string
    - name: latitude
      dtype: float64
    - name: longitude
      dtype: float64
    - name: month
      dtype: int64
    - name: year
      dtype: int64
    - name: move
      dtype: float64
    - name: heading
      dtype: float64
    - name: month_action
      dtype: int64
    - name: year_action
      dtype: int64
  splits:
    - name: train
      num_bytes: 3114758143.367
      num_examples: 72133
    - name: test
      num_bytes: 1380916002
      num_examples: 34000
  download_size: 4229475317
  dataset_size: 4495674145.367001
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*

STRIDE

We develop STRIDE (Spatio-Temporal Road Image Dataset for Exploration), that consists of approximately 82B tokens which were arranged into a total of 6M visual "sentences" or token sequences. The sequences are generated from a relatively small set of 131k panoramic images, along with their metadata and openly available highway system data. Our approach enables a 27x increase in information, allowing for generative world model training. In essence, STRIDE turns real street-view panoramic observations into a navigable, interactive environment suitable for free-play across space and time.

For ease of use and data exploration, we prepared this sample of 10k detokenized paths, which amounts to about 100k projected panoramic images along with its corresponding metadata.

Breakdown

Dataset (STRIDE)

The full tokenized dataset is made available through two downloadable files in a public GCS bucket:

gsutil -m cp gs://tera-tardis/STRIDE-1/training.jsonl . # ~327GB
gsutil -m cp gs://tera-tardis/STRIDE-1/testing.jsonl . # ~9GB

In the future, the fully detokenized dataset will be made available. Should you need it, feel free to contact the authors.

Code (TARDIS)

The code used for training of the model is available on GitHub.

Checkpoints (TARDIS)

The checkpoint/state used for evaluation of the model was saved in MessagePack format and is made available through this downloadable file:

gsutil -m cp gs://tera-tardis/STRIDE-1/checkpoint.msgpack . # ~10GB

Should you need other checkpoints, feel free to contact the authors.

Project Website

The project website is available here.

Contacts

Paper

TARDIS STRIDE: A Spatio-Temporal Road Image Dataset for Exploration and Autonomy