Faiz3005/TartanThermals_Navigability_Dataset

Multi-View Thermal Navigability Dataset

Overview

This dataset contains ultra-low-resolution thermal recordings collected using a wearable multi-camera platform at Carnegie Mellon University in Qatar. It supports research in directional navigability estimation, obstacle detection, and thermal sensor fusion.

The platform uses three Melexis MLX90640 thermal cameras, each capturing 32 × 24 thermal arrays at 4 Hz, covering the left, middle, and right directions.

The dataset includes raw and stitched thermal recordings, automatically generated directional labels, synchronization metadata, supplementary IMU recordings, and thermal homography matrices. RGB imagery used during label generation is not included.

Dataset Contents

The dataset is organized into 11 recording parts, each packaged as a ZIP archive. Each part preserves its original sensor recordings and directory structure.

Thermal Recordings

Thermal data is provided in two formats:

• Unstitched thermal frames: Original individual camera readings stored as JSON files (data_capture_<port>_<frame>_<timestamp>.json). Each camera provides 32 × 24 thermal observations.
• Stitched thermal frames: Combined thermal views stored as NumPy (.npy) files inside stitched_output/. These are generated using homography transformations to align overlapping camera views and contain floating-point thermal data rather than visualized heatmaps.

Synchronization Data

Each recording includes synchronization information in CSV format, associating sensor measurements through timestamps and corresponding capture filenames.

The synchronization CSV contains:

• Anchor information: Reference frame index, stream, timestamp, and filename.
• IMU synchronization: Matched IMU timestamp, time difference from the anchor, and associated file.
• RGB synchronization: Matched RGB timestamp, time difference from the anchor, and associated filename, when available.
• Thermal synchronization: Matched timestamps, time differences, and filenames for individual thermal camera ports.

Missing matches may be represented by placeholder values such as -1. RGB filenames may appear in synchronization records, but the associated RGB imagery is not distributed.

IMU Recordings

Supplementary inertial measurement unit (IMU) recordings are included alongside the thermal and synchronization data.

Directional Labels

The thermal_labels/labels_new.csv file contains automatically generated directional navigability labels associated with the thermal recordings.

Each direction is classified into one of three categories:

• Clear: No detected object overlaps the direction.
• Person: A person detection overlaps the direction.
• Other Object: A non-person object detection overlaps the direction, without a person detection.

Labels were generated using RGB-based object detection and geometric alignment with the thermal camera views.

Homography Matrices

The homography_mats/thermal_homography_mats/ directory contains thermal camera calibration matrices and associated metadata used to geometrically align overlapping thermal camera views.

Directory Structure

├── all_runs_archives/
│   ├── part1.zip
│   ├── part2.zip
│   ├── part3.zip
│   ├── part4.zip
│   ├── part5.zip
│   ├── part8.zip
│   ├── part9.zip
│   ├── part11.zip
│   ├── part12.zip
│   ├── part13.zip
│   ├── part15.zip
│   └── manifest.json
├── thermal_labels/
│   └── labels_new.csv
└── homography_mats/
    └── thermal_homography_mats/

Each recording archive preserves the original sensor directory hierarchy. For example:

part1/
├── thermal-g<session_id>/
│   ├── data_capture_*.json
│   └── stitched_output/
│       └── *.npy
├── sync-g<session_id>/
│   └── [synchronization files]
└── imu-g<session_id>/
    └── [IMU recordings]

The 11 ZIP archives contain 93,057 original files. All archives were verified against their original local files and checked after uploading to Hugging Face.

Usage

Install the Hugging Face Hub library:

pip install huggingface_hub

Download the dataset:

from huggingface_hub import snapshot_download

dataset_path = snapshot_download(
    repo_id="Faiz3005/TartantThermals_Navigability_Dataset",
    repo_type="dataset"
)

Extract the recording archives:

from pathlib import Path
from zipfile import ZipFile

root = Path(dataset_path)
output_dir = Path("all_runs")
output_dir.mkdir(parents=True, exist_ok=True)

for path in sorted((root / "all_runs_archives").glob("part*.zip")):
    with ZipFile(path) as archive:
        archive.extractall(output_dir)

The original part folders and sensor directories are preserved during extraction. Directional labels can be accessed separately at thermal_labels/labels_new.csv.

Thermal Homography Calibration

The supplied thermal homography matrices support geometric alignment and stitching of overlapping thermal camera views.

The thermal calibration tool supports selecting corresponding overlap regions, computing and saving homographies, previewing stitched thermal images, and exporting stitched temperature arrays as .npy files.