enrue1893/Synthetic-Ref-to-Urban-Video-Pairs-Town1-6

Synthetic Reference-to-Urban-Video Pairs (Town 1–6)

Reference-conditioned training data for the Seoul World Model. The model generates an urban driving video — the target — conditioned on a set of reference frames. This repository provides both roles so a complete (reference, target) training pair can be assembled from a single repo.

• 🖼️ Reference — the frames a target is conditioned on. All reference data is the single file references.tar; this is the only reference pool.
• 🎬 Target — the urban driving video sequences the model predicts (RGB + depth + camera), in target_seqs/, one tar per Town × actor.

Captured from the CARLA simulator, Town01–Town06, for both vehicle- and pedestrian-following trajectories. Stored in WebDataset (.tar) format for efficient streaming.

How reference and target link

Every target frame's camera.json carries a matched_references list — the IDs of the reference frames relevant to that frame, accumulated along the trajectory:

{
  "intrinsic":  { ... },
  "extrinsic":  { ... },
  "matched_references": ["Town02/subset_1/0001", "Town02/subset_1/0042", ...]
}

Those IDs index into references.tar, so for any target frame you can look up its conditioning reference frames and build the (reference, target) pair.

Repository structure

Synthetic-Ref-to-Urban-Video-Pairs-Town1-6/
├── references.tar                                  # 151.7 GB — the reference pool (RGB + depth)
│                                                   #   keys: references/Town{01-06}/subset_*/reference/...
└── target_seqs/                                    # ~2.64 TB — target video sequences
    ├── target_seqs_Town01_pedestrian.tar           # ~218 GB
    ├── target_seqs_Town01_vehicle.tar              # ~219 GB
    ├── target_seqs_Town02_pedestrian.tar
    ├── target_seqs_Town02_vehicle.tar
    │   ...                                          # one tar per Town × actor (12 total)
    └── target_seqs_Town06_vehicle.tar

Component	Files	Size
references.tar	1 — shared reference pool	151.7 GB
target_seqs/	12 — Town01–06 × {pedestrian, vehicle}	~2.64 TB
Total		~2.79 TB

Per-frame contents

Each sample (in both references.tar and the target_seqs/ tars) is a WebDataset entry with:

File	Type	Description
*.rgb.png	PIL.Image (1280×704)	RGB image
*.depth.npy	np.ndarray (704, 1280)	Per-pixel depth map
*.camera.json	dict	intrinsic, extrinsic, carla_transform, and (targets) matched_references
*.metadata.json	dict	scene_id, frame_id, town, actor_type

Camera: 1280×704, 90° FOV.

Usage

pip install webdataset huggingface_hub numpy pillow

Stream the reference pool

import io, json
import numpy as np
import webdataset as wds

ref_url = ("https://huggingface.co/datasets/enrue1893/"
           "Synthetic-Ref-to-Urban-Video-Pairs-Town1-6/resolve/main/references.tar")

references = {}  # key -> sample
for s in wds.WebDataset(ref_url).decode("pil"):
    references[s["__key__"]] = {
        "rgb":   s["rgb.png"],
        "depth": np.load(io.BytesIO(s["depth.npy"])),
    }

Stream a target sequence and resolve its references

tgt_url = ("https://huggingface.co/datasets/enrue1893/"
           "Synthetic-Ref-to-Urban-Video-Pairs-Town1-6/resolve/main/"
           "target_seqs/target_seqs_Town01_pedestrian.tar")

for s in wds.WebDataset(tgt_url).decode("pil"):
    rgb      = s["rgb.png"]
    depth    = np.load(io.BytesIO(s["depth.npy"]))
    camera   = json.loads(s["camera.json"])
    ref_ids  = camera.get("matched_references", [])   # -> look up in `references`
    print(s["__key__"], "conditioned on", ref_ids)
    break

The dataset is large (~2.79 TB). Stream shards directly rather than downloading the whole repo, and set HF_TOKEN in your environment if access requires authentication.

Related datasets

Target captures are also published as standalone repositories:

• mkxdxd/carla-dataset — vehicle + pedestrian, Town01–06
• mkxdxd/carla-dataset-ped — pedestrian, 200-frame clips
• mkxdxd/carla-dataset-ped2 — Town05 pedestrian extension
• kaistcvlab/carla-dataset-ped3 — Town05 pedestrian extension

License

Released under CC-BY-4.0. Built using the CARLA simulator (MIT-licensed); see carla.org for simulator/asset terms.