TartanRGBT Dataset

Overview

TartanRGBT is a large-scale, synchronized RGB–Thermal–Depth dataset collected across diverse indoor, outdoor, urban, off-road, and park environments. The dataset is designed to support research in robot perception, visual localization, cross-modal representation learning, thermal vision, and multi-sensor fusion.

All data is organized by day → trajectory (timestamped) → modality, distributed as ZIP archives to enable efficient storage and selective download.

If you use this dataset, please cite our work, AnyThermal: Towards Learning Universal Representations for Thermal Perception, accepted at ICRA 2026. Project website: https://anythermal.github.io/

@misc{maheshwari2026anythermallearninguniversalrepresentations,
      title={AnyThermal: Towards Learning Universal Representations for Thermal Perception},
      author={Parv Maheshwari and Jay Karhade and Yogesh Chawla and Isaiah Adu and Florian Heisen and Andrew Porco and Andrew Jong and Yifei Liu and Santosh Pitla and Sebastian Scherer and Wenshan Wang},
      year={2026},
      eprint={2602.06203},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2602.06203},
}

What's new in this release (10 Hz)

This release upgrades the sampling rate from 1 Hz to 10 Hz — roughly 10× more frames per sequence across all modalities — and fixes the issues reported against the earlier version (see AnyThermal issue #6):

• 10 Hz data for every modality (images, depth, rgb_in_thermal, FFC flags, target timestamps). Frames are now numbered consecutively 00000000, 00000001, 00000002, … (the 1 Hz release used every 10th index …0, …10, …20).
• Fixed TF. The tf/ static transforms inside metadata.zip are no longer identity/zero. They are regenerated from the calibration (calibration/), so the ZED stereo baseline (~0.120 m), the ZED↔IMU extrinsic, and the ZED↔thermal / thermal↔thermal extrinsics are now correct.
• IMU added. metadata.zip now contains zed_imu_raw/ — the full-rate (~100 Hz, not downsampled) ZED IMU stream.
• thermal_left_rect_16.zip added. The 16-bit left thermal images were missing from the previous release and are now shipped.
• Depth and rgb_in_thermal regenerated at 10 Hz with FoundationStereo, using the released-resolution ZED intrinsics.

Trajectory folder names follow the original capture-time naming (undistorted_images_all_cameras_<timestamp>) which matches the layout in custom_datasets/tartanRGBT/splits/sequence.yaml. The RGB-in-thermal modality folder is named rgb_in_thermal/.

All FFC entries, target timestamps, and image frames are aligned at 10 Hz. Per-sequence 10 Hz frame counts are listed in AnyThermal_data_distribution.csv (Number_10Hz).

Sensors & Modalities

Each trajectory contains 10 modality ZIPs + one loose file. Every zip extracts in place under the sequence dir, so the on-disk layout is the same regardless of zip grouping.

Zip / file	Contents	Notes
rgb_in_thermal.zip	RGB (ZED left) images warped into the thermal-left frame	8-bit, FoundationStereo depth reprojection
thermal_left_rect_8.zip	Left thermal, rectified, 8-bit + folded thermal_left_ffc/ (FFC flags)	grayscale
thermal_right_rect_8.zip	Right thermal, rectified, 8-bit + folded thermal_right_ffc/	grayscale
thermal_left_rect_16.zip	Left thermal, rectified, 16-bit	high dynamic range (new in this release)
thermal_right_rect_16.zip	Right thermal, rectified, 16-bit	high dynamic range
zed_left_rect.zip	Left ZED RGB, rectified	960×540
zed_right_rect.zip	Right ZED RGB, rectified	960×540
stereo_depth.zip	Stereo depth maps (metric, meters)	.npy float32, FoundationStereo
odometry.zip	odometry/poses.npy — VO trajectory	own zip (was in metadata)
zed_imu_raw.zip	full-rate (~100 Hz) ZED IMU	own zip (was in metadata)
target_timestamps.txt	loose file — the per-seq master clock	not zipped

All sequences are temporally aligned across modalities and sampled at 10 Hz. Each image modality zip also carries its own timestamps.txt/errors.txt/frames.yaml sidecars (see below). There is no metadata.zip — its contents were redistributed to the modalities they belong to.

target_timestamps.txt — the master clock

One value per frame (seconds), defined by the ZED-left camera at 10 Hz. Line i = the target time of frame i (0000000i.*); every modality was aligned to these times during extraction, so frame index i means the same instant in every folder. The AnyThermal loader uses it to interpolate the odometry onto the frame times.

Per-modality timestamps.txt / errors.txt sidecars (inside each modality zip)

• timestamps.txt — the actual capture time of that modality's matched frame (epoch seconds).
• errors.txt — the alignment error |actual − target| in seconds (~10 ms for thermal, which is not hardware-synced to the ZED). Smaller = tighter sync.

FFC (flat-field correction) — folded into thermal_{left,right}_rect_8.zip

• thermal_left_ffc/data.txt — 1 flag per frame: 0 = good frame, 1 = drop (during FFC recalibration). Already frame-aligned (extractor matched FFC-status by bag time). thermal_left_ffc/errors.txt gives the alignment error. (The FFC Bool topic has no header stamp, so its timestamps.txt was all -1 and is omitted.)

odometry.zip

• odometry/poses.npy — (N, 8) float [timestamp_ns, x, y, z, qx, qy, qz, qw], raw MACVO output at its native rate; the loader interpolates it to target_timestamps.txt.

zed_imu_raw.zip

• data.txt — wx wy wz ax ay az (rad/s, m/s²); orientation.txt — qx qy qz qw; timestamps.txt — sample times (ns), ~100 Hz native rate (not downsampled to 10 Hz).

Calibration

The calibration/ folder is the source of truth for all inter-sensor geometry.

• tf_static.json — the static TF tree (constant across all sequences, so shipped once here instead of per-sequence). Each entry maps parent → child with translation_xyz_m (meters) and quaternion_xyzw. Root frame zed_left_camera_frame. This replaces the earlier per-sequence tf/ (which in the 1 Hz release was a broken identity transform — the subject of issue #6).
• zed.yaml — ZED factory intrinsics/extrinsics at the full 1920×1080 rectified resolution, plus the stereo baseline (tx ≈ 120.022 mm).
• released_zed_rectified_960x540.yaml — intrinsics that match the shipped images. The released ZED images (and therefore stereo_depth and rgb_in_thermal) are 960×540 = the factory 1920×1080 rectified images downsampled by 2×, so fx=369.431, fy=369.4055, cx=484.2525, cy=295.99275 (factory ÷ 2). Metric depth is fx · baseline / disparity with fx=369.431, baseline=0.120022 m.
• thermal_left.yaml, thermal_right.yaml — FLIR 640×512 intrinsics + distortion (equidistant).
• cross_camera_extrinsics.yaml — ZED↔thermal and thermal↔thermal extrinsics.
• zed_imu_extrinsics.json — ZED-left-camera-frame → ZED-IMU-link transform.

Answering issue #6: yes, the released images are downsampled 2× from the factory resolution — use released_zed_rectified_960x540.yaml. The earlier identity/zero TF was a packaging bug; the correct transforms are now in calibration/tf_static.json.

configs/

The exact extraction configuration used to produce this release, for reproducibility:

• configs/tartan_rgbt.yaml — the 10 Hz KITTI-extraction config (synced image/thermal/FFC topics + the full-rate IMU topic).
• configs/handheld.yaml — the corrected static-TF / calibration file the extractor reads (source of calibration/tf_static.json).

Dataset Structure

After download and extraction, the dataset follows the structure below:

TartanRGBT_dataset/
├── day1/
│   └── undistorted_images_all_cameras_<timestamp>/
│       ├── rgb_in_thermal/                          # from rgb_in_thermal.zip
│       │   ├── 00000000_rgb_in_thermal.png
│       │   ├── 00000001_rgb_in_thermal.png
│       │   └── ...
│       ├── thermal_left_rect_8/
│       ├── thermal_right_rect_8/
│       ├── thermal_left_rect_16/
│       ├── thermal_right_rect_16/
│       ├── zed_left_rect/
│       ├── zed_right_rect/
│       ├── stereo_depth/                            # 00000000.npy, 00000001.npy, ...
│       ├── thermal_left_ffc/{data,errors}.txt       # folded into thermal_left_rect_8.zip
│       ├── thermal_right_ffc/{data,errors}.txt      # folded into thermal_right_rect_8.zip
│       ├── zed_imu_raw/{data,orientation,timestamps}.txt   # from zed_imu_raw.zip
│       ├── odometry/poses.npy                        # from odometry.zip
│       ├── target_timestamps.txt                     # loose master-clock file (per seq)
│       ├── (each image folder also has timestamps.txt / errors.txt sidecars)
│       └── [*.zip]                                  # optional: see --delete_zips flag
├── day2/, day3/, day4/, day5/
├── calibration/                                     # incl. tf_static.json (global TF)
├── configs/
├── AnyThermal_data_distribution.csv
├── data_extraction.py
└── README.md

On HuggingFace, only ZIPs + small root files are stored, preserving the same hierarchy: dayX/<trajectory>/*.zip.

Requirements

pip install huggingface_hub

Extraction

The data_extraction.py script downloads every ZIP from this repository and extracts each ZIP in place. After extraction, the layout above is produced.

python data_extraction.py <base_dir> [--delete_zips yes|no]

Arguments

• base_dir — directory where TartanRGBT_dataset/ will be created.
• --delete_zips (default: no) — whether to delete the source ZIPs after extraction succeeds.

Quick sanity check post-extraction

import os, numpy as np
ROOT = "<base_dir>/TartanRGBT_dataset"
seq = "day1/undistorted_images_all_cameras_20250822_115703"
print("rgb frames:    ", len(os.listdir(f"{ROOT}/{seq}/rgb_in_thermal/")))
print("thermal frames:", len(os.listdir(f"{ROOT}/{seq}/thermal_left_rect_8/")))
print("depth frames:  ", len(os.listdir(f"{ROOT}/{seq}/stereo_depth/")))
with open(f"{ROOT}/{seq}/thermal_left_ffc/data.txt") as f:
    ffc = [int(x.strip()) for x in f]
print("ffc length:    ", len(ffc), "  n_dropped:", sum(ffc))
print("ts lines:      ", sum(1 for _ in open(f"{ROOT}/{seq}/target_timestamps.txt")))
poses = np.load(f"{ROOT}/{seq}/odometry/poses.npy")
print("poses shape:   ", poses.shape, "(raw MACVO; interpolated to target_timestamps by the loader)")

You should see rgb frames == thermal frames == depth frames == ffc length == ts lines (all 10 Hz aligned).