anonymous-vision-bench/vision-benchmark

GravCal: Large-Scale Orientation-Diverse Dataset for IMU Gravity Calibration

NeurIPS 2026 Evaluations & Datasets Track

Dataset Description

GravCal is a large-scale dataset specifically designed for single-image IMU gravity calibration. The dataset addresses a critical gap in existing visual-inertial datasets, which exhibit severe upright-pose bias with most frames captured near canonical orientations.

Key Features

• 148,000+ frames with diverse camera orientations
• Explicit coverage of extreme tilts and rotations (0-180°)
• Paired data: RGB image + noisy IMU prior + VIO ground truth
• Real-world IMU noise from Mahony filter integration
• Diverse scenes: Indoor/outdoor with varying lighting conditions
• High-quality labels: VIO-derived gravity with sub-degree accuracy

Dataset Statistics

Property	Value
Total Frames	148,000+
Image Resolution	640×480
Rotation Coverage	0-180° (uniform)
Scene Types	Indoor, Outdoor, Mixed lighting
Train/Val/Test Split	70% / 10% / 20%

Dataset Structure

Data Instances

Each instance contains:

{
    "image_id": "sequence_001_frame_0042",
    "image": PIL.Image,
    "gravity_gt": [0.0234, -0.1234, -0.9922],  # VIO ground truth (3D unit vector)
    "gravity_prior": [0.0456, -0.1456, -0.9856],  # Mahony filter prior (3D unit vector)
    "scene_type": "indoor",  # indoor | outdoor
    "split": "train",  # train | val | test
    "prior_error": 12.3  # Angular error in degrees
}

Data Fields

• image_id: Unique identifier for the frame
• image: RGB image (640×480 JPEG)
• gravity_gt: Ground-truth gravity direction (3D unit vector) from VIO
• gravity_prior: Noisy gravity prior (3D unit vector) from Mahony filter
• scene_type: Scene category (indoor/outdoor)
• split: Data split (train/val/test)
• prior_error: Angular error between prior and ground truth (degrees)

Data Splits

Split	Frames	Percentage
Train	~103,600	70%
Val	~14,800	10%
Test	~29,600	20%

Splits are created by sequence, not random sampling, to prevent data leakage.

Dataset Creation

Source Data

• Hardware: iPhone 12 Pro / 13 Pro Max
• Sensors: Wide camera (12MP) + 6-axis IMU
• Collection: Diverse indoor/outdoor environments with explicit rotation diversity
• Duration: ~150 hours of recording across 300+ sequences

Data Collection

Data was collected with explicit instructions to achieve orientation diversity:

• Systematic rotation sampling
• Coverage of extreme tilts (>60°)
• Various motion patterns (static, walking, running, rotation)
• Multiple lighting conditions (daylight, indoor, low-light)

Annotations

Ground Truth:

• Extracted from ARKit Visual-Inertial Odometry (VIO)
• Validated against public benchmarks (EuRoC, TUM-VI)
• Quality filtering based on tracking confidence

IMU Prior:

• Generated using Mahony filter (Kp=0.5, Ki=0.0)
• Reflects real-world inertial drift
• Error range: 0-90° (concentrated around 10-30°)

Privacy and Ethical Considerations

• Frames containing identifiable individuals are excluded or face-blurred
• Data collected in public/semi-public spaces with appropriate permissions
• No personally identifiable information (PII) included
• Dataset intended for research use only

Benchmark Evaluation

Evaluation Protocols

We provide multiple evaluation settings:

1. In-Domain General: Standard test set
2. Rotation-Stratified: By prior error (0-10°, 10-30°, 30-60°, >60°)
3. Scene-Specific: Indoor / Outdoor / Low-light
4. Cross-Dataset: EuRoC, TUM-VI, UZH-FPV

Baseline Results

Method	Mean Error	Median Error	<10° (%)
IMU Prior (raw)	22.02°	18.45°	42.3%
Image-only	18.76°	15.23°	48.1%
Baseline (ours)	14.24°	11.32°	61.7%

Usage

Loading the Dataset

from datasets import load_dataset

# Load full dataset
dataset = load_dataset("gravcal-neurips2026/gravcal")

# Load only review sample (faster)
dataset = load_dataset("gravcal-neurips2026/gravcal", data_dir="sample")

# Access an instance
sample = dataset["train"][0]
image = sample["image"]
gravity_gt = sample["gravity_gt"]
gravity_prior = sample["gravity_prior"]

Evaluation

import numpy as np

def angular_error(pred, gt):
    """Compute angular error in degrees."""
    cos_angle = np.clip(np.dot(pred, gt), -1.0, 1.0)
    return np.arccos(cos_angle) * 180.0 / np.pi

# Evaluate your method
pred_gravity = model(image, gravity_prior)
error = angular_error(pred_gravity, gravity_gt)

Review Sample

For quick inspection, we provide a representative test sequence:

• Location: sample/ directory
• Size: ~1.5 GB
• Frames: ~1,000 frames
• Coverage: Representative distribution of scenes and rotations

Reviewers can download only the sample for quality inspection without waiting for the full 35GB dataset.

Citation

@inproceedings{gravcal2026,
  title={GravCal: A Large-Scale Orientation-Diverse Dataset and Benchmark for Single-Image IMU Gravity Calibration},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2026}
}

License

• Dataset: CC-BY-4.0
• Code: MIT License

Intended Use

Primary Uses

• Research in visual-inertial perception
• Gravity estimation and IMU calibration
• Mobile AR and robotics applications
• Sensor fusion algorithm development

Prohibited Uses

• Commercial surveillance systems
• Biometric identification or tracking
• Any application violating privacy or ethical guidelines

Contact

For questions, issues, or requests, please open an issue on the code repository.

---

Status: Dataset upload in progress. Full dataset will be available by May 6, 2026.

Version: 1.0.0
Last Updated: May 2026