kaistcvlab/carla-dataset-ped3

CARLA Dataset — Pedestrian (Town05) · Extension 3

A large-scale pedestrian-following driving dataset captured from the CARLA simulator, all in Town05. Provides synchronized RGB + depth + camera parameters along each pedestrian trajectory. Part of the training data for the Seoul World Model. Stored in WebDataset (.tar) format for efficient streaming.

Dataset at a glance

Metric	Value
Town	Town05 only
Actor	Pedestrian
Frames / scene	200
Scenes / shard	2
Total scenes	1,598
Total images	319,600
Shards (.tar)	~800
Total size	~1.49 TB

Captured in two batches, stored as separate subfolders under Town05/:

Subfolder	Scenes	Shards
pedestrian5	799	400
pedestrian6	799	400

Repository structure

carla-dataset-ped3/
└── Town05/
    ├── pedestrian5/
    │   ├── carla-stage2-000000.tar
    │   └── ...
    └── pedestrian6/
        ├── carla-stage2-000000.tar
        └── ...

Each shard holds 2 complete scenes (200 frames each). Sample key: {scene_id}_{frame_idx:03d}.

Per-frame contents

File	Type	Description
*.rgb.png	PIL.Image (1280×704)	RGB image (90° FOV)
*.depth.npy	np.ndarray (704, 1280)	Per-pixel depth map
*.camera.json	dict	intrinsic, extrinsic, carla_transform, matched_references
*.metadata.json	dict	scene_id, frame_id, town, actor_type

The matched_references IDs point into the shared reference pool (references.tar), so each target frame can be paired with its conditioning reference frames.

Usage

pip install webdataset huggingface_hub numpy pillow

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

url = ("https://huggingface.co/datasets/kaistcvlab/carla-dataset-ped3/resolve/main/"
       "Town05/pedestrian5/{carla-stage2-000000..carla-stage2-000010}.tar")

for s in wds.WebDataset(url).decode("pil"):
    rgb      = s["rgb.png"]                            # PIL.Image (1280×704)
    depth    = np.load(io.BytesIO(s["depth.npy"]))    # np.ndarray (704×1280)
    camera   = json.loads(s["camera.json"])
    metadata = json.loads(s["metadata.json"])
    print(s["__key__"], rgb.size, depth.shape, metadata["town"])
    break

Note: the Hugging Face dataset-viewer preview may error on this repo (Cannot write struct type 'target_to_reference_mapping' ...). This affects only the auto-generated Parquet preview — streaming via webdataset is unaffected.

Related datasets

• mkxdxd/carla-dataset — vehicle + pedestrian, Town01–06
• mkxdxd/carla-dataset-ped — pedestrian, 200-frame clips
• mkxdxd/carla-dataset-ped2 — Town05 pedestrian, extension 2
• enrue1893/Synthetic-Ref-to-Urban-Video-Pairs-Town1-6 — reference ↔ target pairs

License

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