Head-inthe-Cloud/ucsc-palms-indoor-localization

Dataset Card for PALMS Indoor Localization Dataset

The PALMS (Plane-based Accessible Indoor Localization Using Mobile Smartphones) dataset is a comprehensive indoor localization dataset collected at the University of California, Santa Cruz (UCSC) campus. The dataset supports both PALMS and PALMS+ algorithms for previous-visit-free indoor localization using commodity mobile devices. It contains RGB images, depth maps, camera poses, floor plans, and ground truth localization data from 80 recording sessions across 4 buildings.

Dataset Details

Dataset Description

The PALMS dataset enables research in indoor visual localization, depth estimation, and floor plan matching. It supports both LiDAR-based (PALMS) and image-based (PALMS+) localization approaches, making it suitable for evaluating various indoor localization algorithms.

• Curated by: Yunqian Cheng, Benjamin Princen, Roberto Manduchi (University of California, Santa Cruz)

• Funded by: National Eye Institute of the National Institutes of Health under award numbers R01EY029260-01 (PALMS) and R01EY036360 (PALMS+)

• Shared by: University of California, Santa Cruz

• Language(s) (NLP): English

• License: CC-BY-NC-4.0 (Non-Commercial Use Only)

Dataset Sources

• Repository: GitHub Repository

• Paper:

  • PALMS: eScholarship (IPIN 2024)
  • PALMS+: arXiv:2511.09724 (WACV 2026)

• Demo: See the GitHub repository for example visualizations and usage scripts

Uses

Direct Use

This dataset is intended for research purposes in:

• Indoor Visual Localization: Training and evaluating algorithms for indoor localization using floor plans
• Depth Estimation: Benchmarking monocular depth estimation models (includes depth maps captured using ARKit LiDAR)
• Point Cloud Reconstruction: Research on 3D point cloud reconstruction from RGB images
• Floor Plan Matching: Developing and testing algorithms that match visual observations to architectural floor plans
• Accessibility Research: Supporting research on accessible indoor navigation systems, particularly for users with visual impairments

The dataset supports both single-shot localization and sequential tracking with particle filters.

Out-of-Scope Use

• Commercial applications: This dataset is licensed for non-commercial use only (CC-BY-NC-4.0)
• Surveillance or tracking: The dataset should not be used for surveillance, tracking individuals, or any privacy-invasive applications
• Real-time navigation systems: While the dataset can inform such systems, it is not intended for direct deployment in production navigation systems without additional validation
• General-purpose computer vision: The dataset is specifically designed for indoor localization tasks and may not be suitable for general computer vision applications

Dataset Structure

The dataset is organized into several main components:

1. main_dataset/

Contains 80 complete recording sessions organized by building:

• BE (Baskin Engineering): 19 sessions
• E2 (Engineering Building 2): 24 sessions
• PS (Physical Sciences Buliding): 18 sessions
• SVC (Silicon Valley Campus): 19 sessions

Each session directory contains:

• images/: RGB images (.png) with AnyLabeling JSON annotations (semantic labels that include transparent surfaces and human)
• poses/: Camera pose files (4×4 transformation matrices) for each frame
• intrinsics/: Camera intrinsic matrices (3×3) for each frame
• depths/: Depth data collected using ARKit LiDAR in JSON format
• dp_depths/: Pre-computed Depth Pro depth maps (.npy files)
• confidences/: Confidence maps for ARKit depth estimation (.png)
• detectedPlanes.json: ARKit detected plane information with timestamps and alignment data
• pano.png: Panoramic image of the scene, collected using iPhone and aligned to the ARKit session reference frame
• label.txt: Ground truth position (x, y coordinates and rotation to map)
• metadata.json: Session metadata
• timeStamps.json: Timestamps for each frame

2. pano_samples/

Subset of sessions with panorama-specific data for quick testing and evaluation:

• 5 sample images per session with corresponding depth maps
• Camera poses and intrinsics
• Panorama metadata (FOV, interval, starting angle)
• Ground truth positions

3. maps/

CSV files containing floor plan geometry for each building:

• BE.csv, E2.csv, PS.csv, SVC.csv: Building floor plans
• Format: Each line contains 4 coordinates (x1, y1, x2, y2) representing wall segments/edges

4. trajectories/

IMU-based trajectory data organized by building and trajectory ID:

• Tracking JSON files: Contains odometry trajectory data in JSON format from either ARKit VIO or RoNIN (IMU-based tracking). Each JSON file contains:

  • starting_vector: Starting position [x, y] in floor plan coordinates
  • ARKit_raw_2D: Raw 2D trajectory data, estimated using ARKit VIO, as an array of [x, y] coordinates over time
  • RoNIN_raw_2D: Raw 2D trajectory data, estimated using the RoNIN model, as an array of [x, y] coordinates over time
  • ARKit_PF: Particle filter processed trajectory as an array of [x, y] coordinates
  • Files are organized by building and trajectory ID (e.g., BE/trajectory_001/tracking_data.json)

• tracking_obs_pairs.csv: Mapping file that pairs tracking trajectories with observation sessions. Contains the following columns:

  • tracking_data_path: Path to the tracking JSON file
  • obs_data_path: Path to the corresponding observation session directory
  • starting_idx: Index in the tracking trace closest to the observation point (used to align tracking sequence with observation)
  • theta: Rotation angle (in radians) applied to align the raw tracking trace with the particle filter reference frame
  • This file enables sequential localization by mapping odometry trajectories to panoramic observation sessions

• Usage: Used for sequential localization with particle filter tracking in test_pp_seq.py. The pairing file allows the system to initialize particle filters at observation points and track user movement using IMU-based odometry between observations.

• For details on pairing process: See utils/prep_trajectories.py for tools to match tracking sequences with observation sessions

Quick Start

To take a quick sense of what the dataset contains, download the example_data.zip file.

Data Loading Tools are provided in the GitHub Repository. Under file utils/file_io.

Dataset Creation

Curation Rationale

The dataset was created to support research on accessible indoor localization systems that do not require prior environmental fingerprinting or external infrastructure. The PALMS system was designed to enhance accessibility for all users, including those with visual impairments, by enabling localization using publicly available floor plans and commodity mobile devices.

Source Data

Data Collection and Processing

Data Collection:

• Data was collected using an iPhone 14 Pro devices with ARKit capabilities at UCSC campus buildings
• RGB images were captured during 360-degree rotational scans at various indoor locations
• ARKit LiDAR depth data was collected simultaneously with RGB images
• Camera poses and intrinsics were recorded for each frame using ARKit
• Ground truth positions were manually annotated with the assistance of ARKit-detected vertical plane data

Data Processing:

• Depth maps were pre-computed using Depth Pro monocular depth estimation model
• Semantic annotations were created using AnyLabeling, we labeled classes like transparent surfaces because monocular depth estimation do not work well with them, and we labeled humans to apply de-identification.
• Panoramic images were captured separately using an iPhone 14 Pro at the label positions. We used the built-in panorama capturing tool in the "camera" app to capture the panorama. Then, we manually undistort and align the panoramas to the images captured in the corresponding scanning sessions.
• Floor plans were digitized from building architectural drawings using the SIM tool, if you are interested in using this tool, please contact Yunqian Cheng at [email protected] or professor Roberto Manduchi at [email protected].

Tools and Libraries:

• ARKit (iOS) for LiDAR data collection
• Depth Pro model for monocular depth estimation
• AnyLabeling for semantic annotation
• Custom Python scripts for data processing and organization

Who are the source data producers?

The dataset was collected and curated by:

• Yunqian Cheng (Primary curator, UCSC)
• Benjamin Princen (Co-curator, UCSC)
• Roberto Manduchi (Principal Investigator, UCSC)
• Loni Halsted-Ruelas (Dataset curation and experimental support, UCSC)

Data was collected at UCSC campus buildings with appropriate permissions.

Annotations

Personal and Sensitive Information

Privacy Measures:

• All human subjects in images have been manually blurred to protect privacy
• Full-body blurring was applied to ensure individuals cannot be identified
• No personally identifiable information (PII) is included in metadata files
• Timestamps do not reveal specific dates that could identify individuals

Data Collection Ethics:

• Data collection was conducted with appropriate institutional permissions
• Privacy considerations were prioritized throughout the dataset creation process
• The dataset complies with privacy and ethical guidelines for research data

Bias, Risks, and Limitations

Limitations

Geographic and Environmental Bias:

• Dataset is limited to UCSC campus buildings, which may not represent all indoor environments
• All data collected in academic/institutional settings, may not generalize to residential or commercial spaces
• Limited to four specific buildings, which may not capture the full diversity of indoor architectural styles

Technical Limitations:

• Depth data quality depends on ARKit LiDAR capabilities and lighting conditions
• Pre-computed depth maps use Depth Pro model, which may have limitations in certain scenarios
• Camera poses may have accumulated errors in texture-less environments
• Floor plans are simplified 2D representations and may not capture all architectural details
• Due to manual labeling, the ground truth poses may have reduced accuracy.

Dataset Size:

• 80 sessions may be limited for large-scale deep learning applications
• Some buildings have fewer sessions than others (imbalanced distribution)

Privacy Considerations:

• While humans are blurred, some contextual information about activities may remain visible

Recommendations

Users should be aware of the following:

1. Generalization: Results obtained on this dataset may not generalize to other indoor environments, particularly those with different architectural styles or lighting conditions

2. Privacy: While privacy measures have been taken, users should be mindful of potential privacy implications when using this dataset

3. Commercial Use: This dataset is licensed for non-commercial use only. For commercial applications, contact the dataset creators

4. Validation: For production systems, additional validation on diverse environments is recommended

5. Ethical Use: The dataset should be used responsibly and in accordance with research ethics guidelines

Citation

BibTeX:

@article{Cheng_Manduchi_2024, title={PALMS: Plane-based Accessible Indoor Localization Using Mobile Smartphones}, author={Cheng, Yunqian and Manduchi, Roberto}, journal={eScholarship, University of California}, url={https://escholarship.org/uc/item/7bw6797s}, year={2024}, month={Aug} }

@article{Cheng_Princen_Manduchi_2025, title={PALMS+: Modular Image-Based Floor Plan Localization Leveraging Depth Foundation Model}, author={Cheng, Yunqian and Princen, Benjamin and Manduchi, Roberto}, journal={arXiv preprint arXiv:2511.09724}, year={2025}, note={Accepted to IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) 2026, Application Track}, url={https://arxiv.org/abs/2511.09724} } APA:

Cheng, Y., & Manduchi, R. (2024). PALMS: Plane-based Accessible Indoor Localization Using Mobile Smartphones. eScholarship, University of California. https://escholarship.org/uc/item/7bw6797s

Cheng, Y., Princen, B., & Manduchi, R. (2025). PALMS+: Modular Image-Based Floor Plan Localization Leveraging Depth Foundation Model. arXiv preprint arXiv:2511.09724. Accepted to IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) 2026, Application Track.

Glossary

• PALMS: Plane-based Accessible Indoor Localization Using Mobile Smartphones
• PALMS+: Extended version of PALMS using monocular depth estimation instead of LiDAR
• CES: Certainly Empty Space - a constraint used for layout matching
• Depth Pro: Foundation model for zero-shot metric monocular depth estimation
• ARKit: Apple's augmented reality framework with LiDAR capabilities
• RoNIN: Model from the paper RoNIN: Robust Neural Inertial Navigation in the Wild: Benchmark, Evaluations, and New Methods
• FOV: Field of View
• BE: Baskin Engineering building
• E2: Engineering 2 building
• PS: Physical Sciences building
• SVC: Silicon Valley Campus

More Information

For more information about using this dataset, please refer to:

• GitHub Repository
• PALMS Paper
• PALMS+ Paper

The repository includes example code, configuration files, and visualization tools for working with the dataset.

Dataset Card Authors

• Yunqian Cheng ([email protected])
• Benjamin Princen
• Roberto Manduchi

Dataset Card Contact

For questions about this dataset, please contact:

• Yunqian Cheng: [email protected]