README.md

---
license: other
task_categories:
- object-detection
- object-detection
---

# 3D Object Detection Dataset

## Overview

- **Data type:** LiDAR (point clouds) + Camera (RGB images)
- **Size:** 245.7 GB, 1,194 ZIP archives
- **Files:** 61,480 PCD + 3,619 JPG
- **Annotations:** 735,483 3D bounding boxes
- **Sessions:** 101 unique recording sessions
- **Camera resolution:** 2880×1860

## Structure

```
.
├── README.md
├── merged_result.parquet         # Main annotation table
├── tf.parquet                    # Transform tree
├── camera_info.parquet           # Camera calibration
├── archive_index.parquet         # Archive lookup index
├── file_catalog.parquet          # File inventory
└── data/
    ├── data_0000.zip
    ├── data_0001.zip
    ├── ...
    └── data_1193.zip
        ├── pcd/{uuid}.pcd        # Point cloud
        └── jpg/{uuid}.jpg        # Camera frame (2880×1860, may be absent)
```

Each archive `data_NNNN.zip` is ≈200 MB. Files inside are from a single temporal segment.

## Tables

### `merged_result.parquet`
Main annotation table (735,483 rows).

| Column | Type | Description |
|---|---|---|
| `frame_id` | `str` | Frame UUID |
| `label` | `str` | Object class (car, pedestrian, cyclist, truck, bus, ...) |
| `data` | `list[float]` | 3D bounding box: [x, y, z, lx, ly, lz, roll, pitch, yaw] |
| `main_timestamp` | `int` | Timestamp in nanoseconds |
| `bag_id` | `str` | Recording session UUID |
| `archive` | `str` | ZIP archive name (`data_NNNN.zip`) |
| `pcd_path` | `str` | Path inside archive (`pcd/{uuid}.pcd`) |
| `jpg_path` | `str` | Path inside archive (`jpg/{uuid}.jpg`) or `null` |

### `tf.parquet`
Transform tree (976,208 rows). Use for converting between coordinate frames.

| Column | Description |
|---|---|
| `timestamp` | Nanoseconds |
| `from_frame`, `to_frame` | Coordinate frames |
| `tx, ty, tz, qx, qy, qz, qw` | Translation + rotation (quaternion) |

### `camera_info.parquet`
Camera calibration (61,480 rows). Use for projecting 3D boxes onto images.

| Column | Description |
|---|---|
| `_timestamp` | Nanoseconds |
| `height`, `width` | 1860×2880 |
| `k` | Intrinsic matrix (3×3) |
| `d` | Distortion coefficients |
| `r` | Rectification matrix |
| `p` | Projection matrix |

### `archive_index.parquet`
Index for O(1) lookup of which archive contains which data.

| Column | Description |
|---|---|
| `archive` | ZIP name |
| `size_mb` | Archive size |
| `file_count`, `pcd_count`, `jpg_count` | File counts |
| `bag_ids` | Session IDs in archive |
| `min_timestamp`, `max_timestamp` | Temporal range |

### `file_catalog.parquet`
Inventory of all files inside the 1,194 ZIP archives (65,496 entries).

| Column | Type | Description |
|---|---|---|
| `archive` | `str` | ZIP archive name |
| `path` | `str` | Full path inside archive (`pcd/{uuid}.pcd` or `jpg/{uuid}.jpg`) |
| `size` | `int` | Uncompressed file size in bytes |
| `csize` | `int` | Compressed size in bytes |



## Usage

```python
import pandas as pd
import zipfile

# Load annotations
df = pd.read_parquet("merged_result.parquet")

# Read a specific point cloud
row = df.iloc[0]
with zipfile.ZipFile(f"data/{row['archive']}") as z:
    pcd_data = z.read(row['pcd_path'])  # bytes → parse as PCD
    if row['jpg_path'] is not None:
        jpg_data = z.read(row['jpg_path'])  # bytes → decode with PIL

# Load transforms
tf = pd.read_parquet("tf.parquet")
# Filter by timestamp frame for coordinate conversion

# Load camera calibration
cam = pd.read_parquet("camera_info.parquet")
```

## Coordinate System

- Bounding boxes are in `base_link` (LiDAR) coordinate frame
- Use `tf.parquet` to convert between `base_link`, `gps`, and other frames
- Camera intrinsics and extrinsics in `camera_info.parquet`

## License

Proprietary. All rights reserved.