HotSight: A Paired RGB-Thermal UAV Dataset for Wildfire Hotspot Segmentation

HotSight is a UAV dataset for mapping combustion states at the pixel level from synchronized RGB and thermal video. It has 951 frame pairs from two prescribed burns. Every RGB pixel is labeled active, smoldering, warm, background, or unknown, and every pair carries a homography that maps the thermal frame into the RGB image. The dataset accompanies a paper under double-blind review. Authorship, citation, and license details will be added after the review.

Where to find what

Data/                       THE DATASET: 951 hand-labeled frame pairs, 12 GB
  Canopy/Canopy-XX/           one folder per clip, see "Inside a clip folder"
  Graminoid/Graminoid-XX/
  samples.csv                 index of all pairs: source videos, frame numbers, timing gap, file paths
  splits_graminoid.csv        evaluation split with Graminoid held out as the test set
  splits_canopy.csv           evaluation split with Canopy held out as the test set
  annotation_label_by_*.csv   class pixel shares per clip and per split

PseudoLabeling/             everything about automatically generated labels, 40 GB
  PseudoDataset10s/           2,214 EXTRA frames with generated labels, 29 GB
  GeneratedAnnotations10s/    generated labels for 775 of the Data/ frames, for comparison with the human labels, 10 GB
  annotation_tool/            the labeling pipeline (YOLO + IR thresholding + SAM), weights, README
  generate_pseudo_dataset.py  the script that produced PseudoDataset10s (needs the source videos)
  GENERATED_LABELS.md         how the generated folders were made and how to use them

html/                       browser tools: video synchronization, manual registration, alignment viewer
scripts/                    the scripts that built the dataset from the raw videos
markdown/                   notes on flights, blended IR clips, removed segments, registration, reproduction
sync.csv                    the 40 source video pairs and their manually estimated IR time offsets
previews/                   downscaled Parquet copies that power the Dataset Viewer at the top of this page. Not for training

I want to...	Get
Train or evaluate a segmentation model	Data/
Look at one clip first	Data/Graminoid/Graminoid-01/ (149 MB, 11 frames, the clip shown below). Clips range from 13 MB to 840 MB
Add automatically labeled training frames	PseudoLabeling/PseudoDataset10s/, read PseudoLabeling/GENERATED_LABELS.md first
Label my own RGB+IR footage	PseudoLabeling/annotation_tool/
Check or refit the RGB-IR alignment	html/viewer.html, html/image-register.html, and the homography/ folder of each clip
Rebuild the dataset from the raw videos	markdown/reproduce.md, scripts/, sync.csv

pip install -U huggingface_hub
hf download hotsightdataset/HotSight --repo-type dataset --local-dir HotSight --include "Data/*"                          # hand-labeled data (12 GB)
hf download hotsightdataset/HotSight --repo-type dataset --local-dir HotSight --include "Data/Graminoid/Graminoid-01/*"   # one clip (149 MB)
hf download hotsightdataset/HotSight --repo-type dataset --local-dir HotSight                                             # everything (59 GB)

The table at the top of this page is the Dataset Viewer. Its config dropdown has two views: default (the 951 hand-labeled pairs) and pseudo (the 2,214 automatically labeled frames, see for example its test split). Both show downscaled previews with the split and per-class pixel percentages, and the filter box accepts expressions such as active_pct > 10.

Inside a clip folder

Data/Graminoid/Graminoid-01/
  rgb/f02400.png              RGB frame, 2688 x 1512
  ir/f02400.png               colorized thermal frame, 640 x 360
  ir/f02400_recovered.png     only in some clips: cleaned thermal frame, use it instead of the original when present
  mask/f02400.txt             the label mask, see "Reading the data"
  mask/f02400_overlay.jpg     labels drawn over the RGB frame, for a quick look
  homography/homography.json  IR -> RGB matrices, one per contiguous frame range
  homography/points.json      the point correspondences the matrices were fitted from

fXXXXX is the frame index in the source RGB video. Hand-labeled frames are multiples of 300, one every 10 s.

What the data looks like

Graminoid-01, frame f02400. Top: the RGB frame and the color IR frame projected into it with the clip's homography, blended equally with the RGB. Bottom: the annotation, one of five classes for every RGB pixel.

Frames Canopy-07/f00300, Canopy-12/f01200, Graminoid-17/f08100, and Graminoid-10/f06300. The thermal camera covers a narrower field of view than the RGB camera, so the projection fills only part of the frame.

The five classes

ID	Class	Operational definition	Overlay color
1	Active	Material with visible flames	dark red (150, 0, 0)
2	Smoldering	Material producing visible smoke but no visible flames	orange (255, 128, 0)
3	Warm	Visibly burned material with no visible flames or smoke	yellow (255, 230, 0)
4	Background	Vegetation or ground with no visible evidence of burning	green (0, 200, 0)
5	Unknown	Smoke, vegetation, or another obstruction prevents a direct assessment of the material	black (0, 0, 0)

Annotators worked from the RGB video and consulted the paired thermal frames as supporting evidence. The labels describe the visible state of the material, not the smoke above it. Warm is a combustion-state label, not a measured temperature. Unknown is a real fifth class that models are expected to predict, not an ignore label.

Reading the data

Each mask is a single line of 2688 x 1512 = 4,064,256 ASCII digits, one per pixel, values 1-5, row-major, no separators.

import numpy as np

def read_mask(path):
    with open(path, "rb") as f:
        return (np.frombuffer(f.read(), dtype=np.uint8) - ord("0")).reshape(1512, 2688)

homography.json maps IR pixels to RGB pixels. A clip may carry several matrices, each for a contiguous frame range. Pick the range that contains the frame, then warp with OpenCV.

import json, cv2, numpy as np

def find_matrix(homography_json, frame):          # frame like "f02400"
    fnum = int(frame.lstrip("f"))
    for s in homography_json["subsets"]:
        lo = int(s["frame_start"].split(".")[0].lstrip("f"))
        hi = int(s["frame_end"].split(".")[0].lstrip("f"))
        if lo <= fnum <= hi:
            return np.array(s["matrix"], dtype=np.float64)
    raise ValueError(f"no homography range covers {frame}")

clip, frame = "Data/Graminoid/Graminoid-01", "f02400"
rgb = cv2.imread(f"{clip}/rgb/{frame}.png")
ir = cv2.imread(f"{clip}/ir/{frame}.png")
H = find_matrix(json.load(open(f"{clip}/homography/homography.json")), frame)
ir_in_rgb = cv2.warpPerspective(ir, H, (rgb.shape[1], rgb.shape[0]))

# The "projected color IR" panels above: equal blend inside the thermal footprint, plain RGB outside it.
inside = cv2.warpPerspective(np.full(ir.shape[:2], 255, np.uint8), H, (rgb.shape[1], rgb.shape[0])) > 127
panel = rgb.copy()
panel[inside] = cv2.addWeighted(ir_in_rgb, 0.5, rgb, 0.5, 0)[inside]

samples.csv always points at the original ir/fXXXXX.png. Where an ir/fXXXXX_recovered.png exists, use that file instead (see "Known quirks").

The two burns

	Canopy	Graminoid
Burn / flights	Mar. 18 / Mar. 18-19, 2026	Mar. 20, 2026
Dominant fuels	One-year rough, wetland fuels, woody debris	Grass, rush, low shrubs, scattered trees
Fire conditions recorded	Active burning on the burn day, residual smoldering and smoke the next day	Active, low-intensity fire
Flight altitude	30-35 m	61 m
Source clip pairs	22	18
Frame pairs	496	455

Both burns took place at Livingston Place, Tall Timbers Research Station and Land Conservancy. A DJI Mavic 2 Enterprise Dual recorded both streams: RGB at 2688 x 1512 and about 30 fps, and an uncooled FLIR Lepton thermal camera (8-14 um) at 640 x 360 and 8.7 fps, upscaled onboard from a 160 x 120 detector. Most thermal clips use the low-gain setting so that embers and residual hotspots do not saturate. Flights mix stationary nadir and oblique views, passes, pauses over smoldering areas, and orbits around hotspots.

Evaluation splits

Each burn unit is held out in turn, so both settings test a change in site, fuel, altitude, and burn day. Within the remaining unit, whole flight blocks (markdown/flights.md) go to validation together, about 10-15% of its frames. Data/splits_graminoid.csv and Data/splits_canopy.csv hold the assignments (frame, split).

Pixel share of each class per split (%):

Class	Graminoid held out: Train	Val	Test	Canopy held out: Train	Val	Test
Active	1.8	1.2	3.7	3.8	3.2	1.7
Smoldering	11.3	3.2	16.2	17.3	6.5	10.1
Warm	12.4	1.4	33.9	37.4	5.3	10.8
Background	72.6	93.6	40.9	37.5	69.6	75.5
Unknown	2.0	0.6	5.2	4.0	15.4	1.8
Frames	427	69	455	406	49	496

Active pixels are under 4% of every split, and the class mix shifts between the units: Graminoid has more warm and smoldering pixels, Canopy more background.

Automatically generated labels

Everything in PseudoLabeling/ comes from the pipeline in PseudoLabeling/annotation_tool/ (a trained YOLO detector and IR hotspot thresholding propose regions, SAM turns them into masks). Two folders hold its output, in the same mask format as Data/.

Folder	Contains	Use it for
PseudoLabeling/PseudoDataset10s/	2,214 new frames from 33 clips, sampled between the hand-labeled ones, with generated labels. pseudo_samples.csv lists them with their anchor frames	Training augmentation. Filter on split first
PseudoLabeling/GeneratedAnnotations10s/	Generated labels for 775 of the existing Data/ frames	Comparison against the human labels only

These labels are generated, not ground truth. The pseudo view of the Dataset Viewer is the quickest way to judge them. Note that the split column of pseudo_samples.csv holds Canopy out as the test site, while the viewer assigns each pseudo frame the split of its anchor frames in the Graminoid-held-out partition, the same as the default view. PseudoLabeling/GENERATED_LABELS.md explains the sampling, the interpolated homographies, and the caveats, and PseudoLabeling/annotation_tool/README.md explains how to run the pipeline on your own footage.

Synchronization and registration

Time. The two cameras record independently. The IR start offset of each of the 40 video pairs was estimated manually with html/sync-video-player.html and is stored in sync.csv (offsets from -1.1 to +0.15 s). Each sampled RGB frame is paired with the temporally nearest thermal frame. The remaining gap has a median of 0.028 s and a 95th percentile of 0.053 s. Thirteen pairs exceed 0.1 s, the largest being 0.798 s. samples.csv records the gap of every pair in nearest_gap_s.

Space. Clips were divided into contiguous frame ranges wherever altitude or viewpoint changed, 122 ranges across the 38 clips with frames. For each range an operator either clicked landmarks visible in both images or adjusted the projected overlay and generated a 6 x 4 grid of correspondences, and an 8-DOF homography was fitted with RANSAC or least squares (html/image-register.html). For the 38 landmark-based fits the control-point RMSE has a median of 2.03 px and a maximum of 15.13 px at full RGB resolution. That is error at the control points only. The timing gap during camera motion, drift within a range, and parallax between the two lenses add error that no single homography removes, which is why the annotations were drawn on the RGB frames directly. html/viewer.html scrubs through any clip with the thermal frame overlaid. Details in markdown/homography.md.

Registration of Graminoid-01, frame f00300. (a, b) Ten correspondences clicked on features visible in both modalities, stored in homography/points.json. (c) The fitted homography warps the IR frame onto the RGB frame. Circles mark the clicked RGB positions and crosses the reprojected IR points.

Known quirks

• Empty clips. Canopy-19 and Graminoid-03 end before the first sampling point, so they have no frames.
• Blended IR. In some clips the camera blended the thermal stream with RGB. The 87 affected frames ship a cleaned ir/fXXXXX_recovered.png next to the original. Use the recovered file when present. Graminoid-17 and -18 are fully colorized and used as is. See markdown/blend.md.
• Odd palette. Canopy-16 uses a thermal color scheme inconsistent with the rest of the dataset.
• Copied homographies. Graminoid-11 and Graminoid-14 have no points.json. Their matrices were copied from the preceding clip of the same flight, as noted in each homography folder.
• Removed segments. Unusable video ranges (transit, camera pointed away from the burn) were cut before sampling. See markdown/junk.md.

Third-party components

PseudoLabeling/annotation_tool/sam_vit_b_01ec64.pth is the Segment Anything ViT-B checkpoint by Meta AI (https://github.com/facebookresearch/segment-anything). html/mp4box.all.min.js is MP4Box.js (https://github.com/gpac/mp4box.js).

Citation

@inproceedings{hotsight2027,
  title     = {HotSight: A Paired RGB-Thermal UAV Dataset and Benchmark for Wildfire Hotspot Segmentation and Localization},
  author    = {Anonymous},
  booktitle = {Under review},
  year      = {2027}
}