autoexpert-cvpr2026-workshop/dataset-public

Auto-Annotation with Expert-Crafted Guidelines: A Study through 3D LiDAR Detection Benchmark

Inspired by the critical bottleneck of data annotation in autonomous driving and recent advancements in foundation models, this dataset introduces a novel evaluation paradigm: Auto-Annotation from Expert-Crafted Guidelines.

Unlike traditional 3D perception benchmarks that rely on massive amounts of annotated 3D point clouds for supervised learning, AutoExpert is designed to evaluate a model's ability to perform multimodal few-shot and zero-shot 3D object detection. Models must comprehend nuanced textual annotator instructions alongside a few 2D visual examples to predict accurate 3D bounding boxes directly in LiDAR data.

✨ Key Characteristics

• Guideline-Driven Learning: Training relies strictly on authentic, expert-crafted textual definitions rather than dense data annotations.
• Zero 3D Training References: Crucially, no 3D LiDAR visual references or 3D bounding boxes are provided during the training phase. The model must bridge the modality gap from 2D/text to 3D space.
• Rigorous Data Cleansing & Re-annotation: We observed that the original 3D annotations in PandaSet often included "ghost" objects—targets visible only in LiDAR but completely occluded or invisible in camera images. To ensure a strict multimodal evaluation, we meticulously re-annotated both the 2D and 3D bounding boxes across the train and val sets, guaranteeing that every annotated object is visually identifiable in the image space.
• Federated 2D Annotations: The few-shot 2D training examples are annotated in a federated manner—only objects belonging to the specific target category are labeled per image, reflecting how real-world guidelines are demonstrated.
• Rich & Long-Tail Vocabulary: Built upon the foundational PandaSet, the dataset challenges models across 25 diverse categories, explicitly testing spatial reasoning on rare, long-tail classes and vulnerable road users (VRUs).

📂 1. Directory Structure

The repository is organized as follows to support the auto-annotation task. It contains the 2D few-shot training examples and the multimodal testing inputs:

cvpr-workshop-challenge-annoexpert-public/
├── train/                        # Few-Shot 2D Examples (Federated Annotation)
│   ├── {image_file}.jpg          # Exemplar images for each category
│   └── metadata.jsonl            # 2D bounding boxes & Expert textual descriptions
├── val/                          # Validation Set
│   ├── {image_file}.jpg          # Multi-view validation images
│   └── metadata.jsonl            # 2D and 3D bounding boxes for validation frames
├── test/                         # Multimodal Testing Set (Inputs only)
│   ├── {image_file}.jpg          # Multi-view test images (200 target frames)
│   └── metadata.jsonl            # Image metadata and linking indices
├── .gitattributes                # Git LFS configuration
└── README.md                     # Dataset documentation

(Note: Raw LiDAR point clouds and calibration data are hosted in a separate repository to optimize size. See Section 4 for loading instructions).

📄 2. Task Formulation & Data Formats

2.1 The Unified metadata.jsonl Schema

Instead of parsing separate annotation text files, participants can read all multi-modal context directly from metadata.jsonl. Each record follows this unified schema:

{
  "file_name": "Car&001_front_camera_02.jpg",
  "target_category": "Car",
  "expert_description": "Sedans, coupes, SUVs.",
  "seq_id": "001",
  "frame_idx": "02",
  "camera_name": "front_camera",
  "objects": {
    "bbox": [[10.5, 20.2, 50.0, 60.0]], 
    "categories": ["Car"]
  },
  "3d_boxes": "[...]" 
}

2.2 Expert Guidelines & Few-Shot Examples (Training)

Participants must rely on the provided guidelines and few-shot examples to understand the 25 target categories.

• Expert Descriptions: The official annotator definitions are embedded directly in the expert_description field of train/metadata.jsonl.
• Federated & Cleansed 2D Annotations: The objects field contains 2D bounding boxes ([x_min, y_min, width, height]). In a given training image, only objects belonging to the target_category are annotated, while objects of other classes are intentionally ignored. All annotated objects are verified to be visually identifiable.

2.3 Validation Set (val/)

To help validate models before submitting to the evaluation server, a validation set on 8 specific keyframes is provided.

• 2D Annotations: Comprehensive 2D bounding box annotations for all visible objects are provided in the objects field of val/metadata.jsonl. These are specifically provided to help participants select and validate their 2D detection models.
• Re-annotated 3D Ground Truth: The 3d_boxes field contains the ground truth 3D bounding boxes. Unlike the original PandaSet, these boxes have been manually cleansed and re-annotated to ensure strong image-LiDAR alignment (removing fully occluded objects). You can use this set to locally evaluate your model's 3D detection metrics (mAP, NDS).

2.4 Test Sensor Data (Evaluation)

The evaluation focuses on 192 specific keyframes in the test set.

• Inputs Only: The test/metadata.jsonl contains image metadata and linking indices, but the objects and 3d_boxes fields are empty (or contain empty placeholders to maintain schema consistency).
• LiDAR Data Access: You must download the separate LiDAR sequence repository. You can then use the seq_id, frame_idx, and camera_name from the metadata to load point clouds and sensor calibration via the PandaSet Devkit.

📝 3. Submission Format

For the evaluation server to process your predictions, you must submit your 3D detection results in a highly specific JSON format.

Participants must generate a single submission.json file containing all predictions for the 200 test images. The JSON file should contain a list of dictionaries, where each dictionary represents a single predicted 3D bounding box.

JSON Structure Example:

[
  {
    "seq_id": "001",
    "frame_idx": 29,
    "frame_token": "001_front_left_camera_000029",
    "label": "Car",
    "score": 0.8,
    "box_3d": [10.5, -3.2, -1.0, 4.5, 1.8, 1.5, 0.12]
  },
  {
    "seq_id": "001",
    "frame_idx": 29,
    "frame_token": "001_front_left_camera_000029",
    "label": "Pedestrian",
    "score": 0.9,
    "box_3d": [12.1, -1.5, -0.8, 0.5, 0.6, 1.7, 0.05]
  }
]

Field Definitions:

• seq_id (String): The sequence identifier from the PandaSet dataset (e.g., "001").
• frame_idx (Integer): The frame index within the sequence (e.g., 29).
• frame_token (String): The unique identifier for the specific camera frame, formatted as {seq_id}{camera_name}{frame_idx}.
• label (String): The predicted category name. It must exactly match one of the 25 official classes.
• score (Float): The confidence score of the prediction (between 0.0 and 1.0).
• box_3d (List of Floats): The 3D bounding box parameters in the LiDAR coordinate system. It must contain exactly 7 values: [x, y, z, l, w, h, yaw].
  • x, y, z: The 3D center coordinates.
  • l, w, h: The length, width, and height of the box.
  • yaw: The orientation/yaw angle in radians.

Please ensure your final submission is a single valid JSON file named submission.json and upload it to our official evaluation server at https://huggingface.co/spaces/autoexpert-cvpr2026-workshop/auto3D.

🏷️ 4. Class Categories

The dataset includes 25 distinct object categories, requiring models to handle diverse traffic participants and nuanced object definitions.

Vehicle & Transport	Vulnerable Road Users (VRU)	Infrastructure & Obstacles
Car	Pedestrian	Temporary_Construction_Barriers
Pickup_Truck	Pedestrian_with_Object	Cones
Medium-sized_Truck	Bicycle	Signs
Semi-truck	Motorcycle	Rolling_Containers
Bus	Motorized_Scooter	Pylons
Tram_or_Subway	Personal_Mobility_Device	Road_Barriers
Emergency_Vehicle	Animals-Other	Construction_Signs
Other_Vehicle-Construction_Vehicle		Towed_Object
Other_Vehicle-Uncommon
Other_Vehicle-Pedicab

Note: Participants must adhere to the specific definitions for each class as outlined in the expert descriptions.

🚀 5. Getting Started

We recommend using the Hugging Face datasets library to load the images and metadata, and the official pandaset devkit to load the physical sensor data.

5.1 Environment Setup

# 1. Install required libraries
git clone git@github.com:scaleapi/pandaset-devkit.git
cd pandaset-devkit/python
pip install .

# 2. Clone the external LiDAR sequence repository (Ensure Git LFS is installed)
git lfs install
git clone https://huggingface.co/datasets/autoexpert-cvpr2026-workshop/seq ./seq

5.2 Data Loading Example (Python)

The following code demonstrates how to load images, 2D/3D annotations, and corresponding LiDAR point clouds and sensor calibration across different splits.

import json
from datasets import load_dataset
from pandaset import DataSet

# 1. Load the HF dataset (Images + Metadata)
hf_dataset = load_dataset("autoexpert-cvpr2026-workshop/dataset-public")

# 2. Initialize the PandaSet Devkit with the cloned seq_data path
panda_data = DataSet('./seq')

# ==========================================
# [TRAIN SPLIT]: Images and 2D Annotations
# ==========================================
train_sample = hf_dataset['train'][0]
image = train_sample['image']                             # PIL Image
target_category = train_sample['target_category']         # e.g., "Car"
expert_desc = train_sample['expert_description']          # Textual guideline
bbox_2d = train_sample['objects']['bbox']                 # Federated 2D bbox

print(f"Train - Class: {target_category} | Guideline: {expert_desc}")

# ==========================================
# [VAL SPLIT]: Multimodal + 3D Ground Truth
# ==========================================
val_sample = hf_dataset['val'][0]
seq_id = val_sample['seq_id']                             # e.g., "001"
frame_idx = int(val_sample['frame_idx'])                  # e.g., 2
camera_name = val_sample['camera_name']                   # e.g., "front_camera"

# Load corresponding LiDAR & Calibration via PandaSet Devkit
seq = panda_data[seq_id]
seq.load_lidar()
seq.load_camera(camera_name)

# Extract multi-modal sensor data
point_cloud = seq.lidar[frame_idx]                        # Raw LiDAR points
intrinsics = seq.camera[camera_name].intrinsics           # Camera intrinsics
extrinsics = seq.camera[camera_name].poses[frame_idx]     # Sensor poses/extrinsics

# Extract annotations
val_image = val_sample['image']
val_bbox_2d = val_sample['objects']['bbox']               # Comprehensive 2D bbox
val_bbox_3d = json.loads(val_sample['3d_boxes'])          # 3D Ground Truth

print(f"Val - Loaded Point Cloud shape: {point_cloud.shape}")

# ==========================================
# [TEST SPLIT]: Multimodal Inputs Only
# ==========================================
test_sample = hf_dataset['test'][0]
seq_id = test_sample['seq_id']
frame_idx = int(test_sample['frame_idx'])
camera_name = test_sample['camera_name']

# Load corresponding LiDAR & Calibration via PandaSet Devkit
seq = panda_data[seq_id]
seq.load_lidar()
seq.load_camera(camera_name)

# Model Inputs
test_image = test_sample['image']
point_cloud = seq.lidar[frame_idx]
intrinsics = seq.camera[camera_name].intrinsics
extrinsics = seq.camera[camera_name].poses[frame_idx]

# Model should predict 3D bounding boxes based on the above inputs!

⚖️ 6. License & Citation

This dataset is built upon PandaSet and is subject to the PandaSet License Terms.

(Note: The full citation for the AutoExpert baseline paper will be updated here upon the conclusion of the double-blind review process. Below is the anonymous placeholder.)

@misc{ma2025auto,
  title = {Auto-Annotation with Expert-Crafted Guidelines: A Study through 3D LiDAR Detection Benchmark},
  author = {Ma, Yechi and Hua, Wei and Kong, Shu},
  booktitle = {arXiv:2506.02914},
  year = {2025}
}

@inproceedings{xiao2021pandaset,
  title = {PandaSet: Advanced Sensor Suite Dataset for Autonomous Driving},
  author = {Xiao, Peng and Shao, Zili and Hao, Shaoyu and others},
  booktitle = {IEEE International Intelligent Transportation Systems Conference (ITSC)},
  year = {2021}
}