mani-ramanagopal/subtle-heat-flows

Subtle Heat Flows

Raw microbolometric video data accompanying "Revealing Subtle Heat Flows all around us using Microbolometric Videos" (IEEE International Conference on Computational Photography, ICCP 2026).

• Paper: PDF
• Project page: https://revealing-subtle-heat.github.io/
• Poster: PDF
• Data capture code: GitHub
• Data processing code: GitHub
• Authors: Mani Ramanagopal, Akihiko Oharazawa, Sriram Narayanan, Zeqing Yuan, Srinivasa Narasimhan (Carnegie Mellon University)

Dataset Summary

Low-cost microbolometric (thermal) cameras are widely assumed to be too noisy to capture subtle heat transport. This dataset provides the raw, unprocessed video captures used in the paper to demonstrate that a simple per-pixel 1D temporal transform followed by spatial denoising in the coefficient domain reveals rich spatiotemporal heat flow phenomena from ordinary low-cost hardware.

The dataset contains 25 raw capture sequences across 9 distinct scenes, two of which include parameter sweeps (fan speed, illumination level).

Access

This dataset is gated: you must accept the access conditions (to share contact information) before downloading. This is used only to track usage of the dataset and does not reflect any privacy concern with the captured content.

Dataset Structure

Directory layout

Each scene lives in its own folder containing a single compressed archive:

<scene_id>_<scene_name>/
└── <scene_id>_<scene_name>.npz.zst

25 folders in total. Scene IDs beginning with D are dual-camera captures; the single ID beginning with S is a single-camera capture (see "Camera configurations" below). Some scenes could have dropped frames.

Scene ID	Name	Notes	Length
D01	bunny_wind_draft	3D printed bunny, convection from gentle breeze	1201
D02	plain_wall	Reference/calibration scene	1200
D03	bunny_light_right	3D printed bunny, illumination from the right	1201
D04	bunny_light_left	3D printed bunny, illumination from the left	1201
D05	vehicle_exhaust	outdoor road scene, vehicle exhaust and radiator airflow revealed	1202
D06	bulb_bandpass_filters	LWIR bandpass filters, spectral emission	1801
D07	yoda_eyes	Thermal reflection of two humans in the background	1200
D08	book_colorchart_L0–L5	Book/colorchart heating up due to light absorption, 6 illumination levels	variable (1178 < T < 1211)
D09	cardboard_fan_L0–L10	Black cardboard heated by incandescent bulb is cooled by a PC fan	variable (1187 < T < 1213)
S01	bowl_polarizer	Manually rotated wire-grid polarizer in LWIR, single-camera	1202

Each .npz.zst is a zstandard-compressed NumPy .npz archive.

Camera configurations

All data was captured at 60 fps with two identical FLIR Boson+ (22640A024-6IAAX) LWIR microbolometer cameras.

• D-prefixed scenes (dual-camera): two synchronized cameras capture the same scene simultaneously, differing in whether the camera's internal filter is engaged.
  • _A suffix → internal filter off
  • _B suffix → internal filter on
• S-prefixed scenes (single-camera): one camera only, internal filter off (no _A/_B suffix).

Archive contents

Dual-camera (D*) scenes — each .npz contains 6 arrays:

Key	Shape	dtype	Description
raw_thr_frames_A	(T, 514, 640, 1)	uint16	Raw thermal frames, filter off. First 2 rows of each frame are camera telemetry, not image data.
raw_thr_tstamps_A	(T,)	float64	Per-frame timestamp, filter off, already adjusted to PC epoch time (offset included).
thr_cam_timestamp_offset_A	scalar/(1,)	float64	Offset added to the camera's power-on-referenced hardware timestamp to map it to PC epoch time.
raw_thr_frames_B	(T, 514, 640, 1)	uint16	Same as above, filter on.
raw_thr_tstamps_B	(T,)	float64	Same as above, filter on.
thr_cam_timestamp_offset_B	scalar/(1,)	float64	Same as above, filter on.

Single-camera (S*) scenes — each .npz contains the same 3 arrays without the _A/_B suffix:

Key	Shape	dtype	Description
raw_thr_frames	(T, 514, 640, 1)	uint16	Raw thermal frames, filter off. First 2 rows are telemetry.
raw_thr_tstamps	(T,)	float64	Per-frame timestamp, already adjusted to PC epoch time.
thr_cam_timestamp_offset	scalar/(1,)	float64	Offset added to map camera hardware time to PC epoch time.

Notes:

• Frame height is 514 rows because the top 2 rows are FLIR Boson+ telemetry data, not part of the 512-row thermal image; crop these before visualizing or processing as an image.
• raw_thr_tstamps* already has thr_cam_timestamp_offset* applied — the offset field is provided for reference/traceability rather than required for use.
• D09_cardboard_fan_L0–L10: L indexes fan speed level (11 levels).
• D08_book_colorchart_L0–L5: L indexes illumination level (6 levels).

Loading example

import numpy as np
import zstandard as zstd
import io

path = "D01_bunny_wind_draft/D01_bunny_wind_draft.npz.zst"
with open(path, "rb") as f:
    decompressed = zstd.ZstdDecompressor().decompress(f.read())
data = np.load(io.BytesIO(decompressed))

frames_A = data["raw_thr_frames_A"]           # (T, 514, 640, 1) uint16
tstamps_A = data["raw_thr_tstamps_A"]         # (T,) float64, PC epoch time
image_A = frames_A[:, 2:, :, 0]                # drop 2 telemetry rows -> (T, 512, 640)

Dataset Size

Total size: 24.7 GB across 25 files.

Citation

@inproceedings{ramanagopal2026revealingsubtleheat,
  title     = {Revealing Subtle Heat Flows all around us using Microbolometric Videos},
  author    = {Ramanagopal, Mani and Oharazawa, Akihiko and Narayanan, Sriram and
               Yuan, Zeqing and Narasimhan, Srinivasa},
  booktitle = {IEEE International Conference on Computational Photography (ICCP)},
  year      = {2026}
}

License

This dataset is released under the CC BY 4.0 license.