mkxdxd/carla-dataset

CARLA Dataset (Target)

A large-scale driving dataset captured from CARLA simulator, containing RGB images and depth maps with camera parameters for autonomous driving research.

Tar File Structure

This dataset is stored in WebDataset format for efficient streaming and loading.

Sharding Strategy

• 3 scenes per shard: Each .tar file contains exactly 3 complete scenes
• 93 frames per scene: Each scene is a video sequence of 93 consecutive frames
• ~279 frames per tar: 3 scenes × 93 frames = 279 frames per shard

Repository Structure

carla-dataset/
├── Town01/
│   ├── vehicle/
│   │   ├── carla-stage2-000000.tar  ← Scenes 1-3
│   │   ├── carla-stage2-000001.tar  ← Scenes 4-6
│   │   └── ...
│   └── pedestrian/
│       └── ...
├── Town02/
│   └── ...
├── Town03/
│   └── ...
├── Town04/
│   └── ...
├── Town05/
│   └── ...
└── Town06/
    └── ...

Shard Contents (WebDataset format)

Each tar file contains samples with the following files per frame:

{scene_id}_{frame_idx:03d}.rgb.png        ← RGB image (1280×704)
{scene_id}_{frame_idx:03d}.depth.npy      ← Depth map (numpy array, 704×1280)
{scene_id}_{frame_idx:03d}.camera.json    ← Camera parameters + matched reference frames
{scene_id}_{frame_idx:03d}.metadata.json  ← Scene info (scene_id, frame_id, town, actor_type)

Key Format: {scene_id}_{frame_idx:03d} (e.g., scene_001_000, scene_001_001, ..., scene_001_092)

Data Format

Each sample contains:

• rgb: PIL.Image (1280×704) - RGB image
• depth: np.ndarray (704, 1280) - Depth map
• camera: dict containing:
  • intrinsic: Camera intrinsic matrix
  • extrinsic: Camera extrinsic matrix
  • matched_references: List of reference image IDs accumulated along the trajectory
• metadata: dict containing:
  • scene_id: Unique scene identifier
  • frame_id: Frame index within the scene (0-92)
  • town: CARLA town name (Town01-Town06)
  • actor_type: Type of actor being followed ("vehicle" or "pedestrian")

---

Camera JSON with Reference Frame Mapping

Each camera.json contains camera parameters and matched reference frames:

{
  "intrinsic": { ... },
  "extrinsic": { ... },
  "carla_transform": { ... },
  "matched_references": ["subset_0/0001", "subset_0/0042", "subset_0/0083"]
}

Reference Accumulation: As the camera moves through the scene, reference frames are accumulated:

• Frame 0: [ref_A] - first reference encountered
• Frame 20: [ref_A, ref_B] - new reference added
• Frame 40: [ref_A, ref_B, ref_C] - another reference added
• ...

This allows each target frame to know which reference images are relevant based on position and viewing angle.

---

Dataset Statistics

Dataset Summary

Town	Mode	Scenes	Images	Avg Images/Scene
Town01	pedestrian	520	48,360	93.0
Town01	vehicle	490	45,570	93.0
Town01	TOTAL	1,010	93,930	93.0
Town02	pedestrian	509	47,337	93.0
Town02	vehicle	500	46,500	93.0
Town02	TOTAL	1,009	93,837	93.0
Town03	pedestrian	500	46,500	93.0
Town03	vehicle	500	46,500	93.0
Town03	TOTAL	1,000	93,000	93.0
Town04	pedestrian	500	46,500	93.0
Town04	vehicle	530	49,290	93.0
Town04	TOTAL	1,030	95,790	93.0
Town05	pedestrian	500	46,500	93.0
Town05	vehicle	500	46,500	93.0
Town05	TOTAL	1,000	93,000	93.0
Town06	pedestrian	500	46,500	93.0
Town06	vehicle	500	46,500	93.0
Town06	TOTAL	1,000	93,000	93.0

Grand Total

Metric	Value
Total Scenes	6,049
Total Images	562,557
Towns	6 (Town01-06)

---

Example Usage

📝 Full example code is available in example_usage.py

Installation

pip install datasets torch pillow numpy webdataset

Basic Usage with webdataset

import webdataset as wds
from huggingface_hub import hf_hub_url
import json
import numpy as np
from PIL import Image
import io

# Stream dataset from HuggingFace
url = "https://huggingface.co/datasets/mkxdxd/carla-dataset/resolve/main/Town01/pedestrian/{carla-stage2-000000..carla-stage2-000010}.tar"

dataset = wds.WebDataset(url).decode("pil")

for sample in dataset:
    key = sample["__key__"]
    rgb = sample["rgb.png"]           # PIL.Image (1280x704)
    depth = np.load(io.BytesIO(sample["depth.npy"]))  # numpy array (704x1280)
    camera = json.loads(sample["camera.json"])        # dict
    metadata = json.loads(sample["metadata.json"])    # dict
    
    print(f"Key: {key}")
    print(f"RGB size: {rgb.size}")
    print(f"Depth shape: {depth.shape}")
    print(f"Scene: {metadata['scene_id']}, Town: {metadata['town']}")
    break

With PyTorch DataLoader

import webdataset as wds
import torch
from torch.utils.data import DataLoader
import json
import numpy as np
from PIL import Image
import io

def decode_sample(sample):
    return {
        "rgb": sample["rgb.png"],
        "depth": np.load(io.BytesIO(sample["depth.npy"])),
        "camera": json.loads(sample["camera.json"]),
        "metadata": json.loads(sample["metadata.json"]),
    }

url = "https://huggingface.co/datasets/mkxdxd/carla-dataset/resolve/main/Town01/pedestrian/{carla-stage2-000000..carla-stage2-000010}.tar"

dataset = (
    wds.WebDataset(url)
    .decode("pil")
    .map(decode_sample)
)

dataloader = DataLoader(dataset, batch_size=8, num_workers=4)

for batch in dataloader:
    rgb_batch = batch["rgb"]
    depth_batch = batch["depth"]
    # ... training code
    break

Dataset generated from CARLA Simulator