shadman-sakib/MPSC-RespView

Overview · Contents · Protocol · Statistics · Quick start · Benchmark · Privacy · Access · Cite · Contact

Cite this dataset · WACV 2027

Pixels-to-Phase: Reliability-Aware Latent State-Space Modeling for Contactless Respiratory Estimation

Shadman Sakib and Nirmalya Roy · IEEE/CVF Winter Conference on Applications of Computer Vision (WACV 2027, Main Track)

Paper website  ·  BibTeX

Overview

MPSC-RespView is a collection of 73 RGB video recordings of human breathing, captured indoors at close range and outdoors at long range, for developing and evaluating contactless respiratory-rate and respiratory-waveform estimation. Indoor recordings show 15 subjects from several camera views and postures: frontal, back, top-angular, lying and side. Outdoor recordings vary the subject–camera distance (4–13 m) and include normal breathing, breath holding and forced breathing.

It is the In-House Dataset v1.0 used in Pixels-to-Phase: Reliability-Aware Latent State-Space Modeling for Contactless Respiratory Estimation (WACV 2027), and is released here for the research community. See the paper website for the method, results and supplementary material.

Recordings	73 (50 indoor · 23 outdoor)
Indoor	15 subjects · mostly 1920×1078 · 30 fps · 1.6–4.8 min per recording
Indoor views	frontal · back · top-angular · lying · side
Outdoor	5 subjects · 4–13 m · 1920×1078 · 30 fps · about 56–59 s per recording
Total	2.72 hours of video · 13.5 GB
Sensor	Standard RGB video cameras · no audio · no depth
Privacy	Faces anonymized · no audio · coded subject IDs · no names or other identifiers
Paper	Pixels-to-Phase (WACV 2027)
Access	Open download · License: CC BY-NC 4.0 · Terms of use apply

Why this dataset

• Viewpoint and posture diversity. Frontal, back, top-angular, lying and side views, so respiratory motion appears at the chest, abdomen, shoulders and clothing folds rather than in one fixed geometry.
• Long-range recordings. Outdoor clips in which the torso covers few pixels, where the visual signal is weak and the failure modes of existing methods show up.
• Hard conditions on purpose. Spontaneous and paced breathing (slow, regular, fast), breath holds and forced breathing, so models are tested on rate changes and on intervals with no breathing motion.
• Subject-wise by design. Every recording carries a coded subject ID, so subject-wise train/test splits are straightforward.

What's inside

MPSC-RespView/
├── README.md
├── LICENSE
├── DATA_USE_AGREEMENT.md
├── CITATION.cff
├── CHANGELOG.md
├── metadata.csv                 # one row per recording (see schema below)
├── SHA256SUMS.txt               # integrity checksums for every video
├── indoor/
│   ├── sub_01/                # e.g. sub_01_back.mp4
│   ├── sub_02/
│   └── …                        # 15 subjects
├── outdoor/
│   ├── out_01/                # e.g. out_01_04m.mp4 (distance in metres)
│   └── …                        # 5 subjects
├── assets/
└── scripts/                     # loader example and checksum verification

Download only what you need. The indoor videos are large, so use allow_patterns (see Quick start).

Recording protocol

Indoor. Participants were guided through breathing exercises (natural, slow, regular and fast) in several postures and facing directions, and were recorded from multiple angles with standard RGB cameras. Each subject has up to four recordings, one per view. Respiration-belt reference signals were recorded during the sessions; they are kept private and are not part of this release. The acquisition protocol follows the one used for the Sleep benchmark.

Outdoor. One-minute recordings at several subject–camera distances, each with roughly 20 s of normal breathing, 20 s of breath holding and 20 s of forced breathing. Outdoor subject IDs (out_01 …) are independent of the indoor IDs, and no file name encodes who the participant is.

Ethics. Participants were healthy adults who gave informed consent under a UMBC Institutional Review Board (IRB)-approved protocol. No audio was recorded. Videos carry only a coded study ID.

File naming

Pattern	Meaning
indoor/sub_XX/sub_XX_<view>.mp4	Indoor recording of subject XX in the given view
outdoor/out_YY/out_YY_<DD>m.mp4	Outdoor recording of outdoor subject YY at DD metres

Indoor views:

View	Posture	Description	Recordings
frontal	seated	Seated, facing the camera (close to medium shot of the torso)	12
back	seated	Seated, back to the camera	13
top_angular	seated	Seated, camera raised and angled down, wider shot	10
lying	lying	Lying down (on the floor; on a sofa for subjects 11–13)	13
side	seated	Seated, profile (side) view, portrait video	2

metadata.csv is the authoritative index. A second recording of the same view for one subject gets a _2 suffix.

Statistics

metadata.csv schema

Column	Description
file_name	Path of the video, relative to the repository root
domain	indoor or outdoor
subject_id	Coded subject ID (sub_XX indoor, out_YY outdoor)
view	Indoor camera view: frontal, back, top_angular, lying, side (blank for outdoor)
posture	seated or lying
distance_m	Subject–camera distance in metres (outdoor only)
width, height, orientation, fps	Video resolution, portrait or landscape, and frame rate
duration_s, frames	Length of the recording
codec, bitrate_mbps, size_mb	Encoding details and file size
original_name	Internal recording name (indoor only), for matching the paper's split manifests
sha256	SHA-256 checksum of the video

Video formats

The recordings come from several cameras, so resolution is not uniform. Check width, height and orientation in metadata.csv and resize or crop in your data loader.

Resolution (W×H)	Codec	Indoor	Outdoor
1920×1078	h264	39	23
1440×1080	h264	6	–
640×480	mpeg4	2	–
1080×1920 (portrait)	h264	2	–
480×848 (portrait)	h264	1	–

Quick start

1. Install the dependencies:

pip install -U huggingface_hub pandas opencv-python

2. Download the index and the part you need:

from huggingface_hub import snapshot_download
import pandas as pd, cv2

repo = "shadman-sakib/MPSC-RespView"

# Index + the outdoor set (small). Add e.g. "indoor/sub_01/*" for indoor subjects.
root = snapshot_download(repo, repo_type="dataset",
                         allow_patterns=["metadata.csv", "outdoor/*", "scripts/*"])
meta = pd.read_csv(f"{root}/metadata.csv")
print(meta.groupby("domain").size())

# Read one video
row = meta.query("domain == 'outdoor' and distance_m == 7").iloc[0]
cap = cv2.VideoCapture(f"{root}/{row.file_name}")
fps = cap.get(cv2.CAP_PROP_FPS)
ok, frame = cap.read()
print(row.file_name, frame.shape, fps)

Select recordings by view:

lying = meta.query("domain == 'indoor' and view == 'lying'")
print(lying[["file_name", "subject_id", "duration_s"]].head())

4. Verify your download (optional but recommended):

python scripts/verify_checksums.py --root <download_dir>

Benchmark

Results from the paper on this dataset (all 73 recordings, 5-fold subject-wise evaluation, computed against respiration-belt references that are not distributed). MAE and RMSE are in breaths per minute; R is the Pearson correlation of the estimated and reference waveforms. Split manifests will be released with the code.

Within-dataset (selected methods from Table 1 of the paper):

Method	MAE ↓	RMSE ↓	R ↑
POS	3.26	4.62	0.69
Green	4.07	5.43	0.59
CHROM	3.33	4.98	0.66
ICA	3.72	5.14	0.63
Intensity + Peak Counting	4.10	5.79	0.60
PhysNet	2.32	3.21	0.79
PhysFormer	1.97	3.04	0.82
EfficientPhys	2.13	3.38	0.80
SpikingPhys	2.08	2.91	0.82
AIRFlowNet	1.82	3.02	0.81
RRPIPS	1.75	2.47	0.86
FLOW	1.36	2.22	0.89
Pixels-to-Phase	1.28	2.06	0.90

Held-out (trained without this dataset, tested on it):

Method	MAE ↓	RMSE ↓	R ↑
PhysNet	3.18	4.06	0.71
PhysFormer	2.39	3.79	0.76
EfficientPhys	2.84	4.31	0.72
SpikingPhys	2.52	3.36	0.80
AIRFlowNet	2.05	3.44	0.78
RRPIPS	2.08	2.81	0.84
FLOW	1.61	2.55	0.87
Pixels-to-Phase	1.38	2.04	0.90

Privacy and ethics

• Consent. All participants gave informed consent under a UMBC IRB-approved protocol, including consent to record their image and to share anonymized data with other researchers.
• Anonymization. Faces are blurred in every frame before release. Videos contain no audio and no names; subjects are identified only by a coded ID. File metadata was checked for identifying tags.
• No re-identification. You may not try to identify any participant, link the data to other records, or attempt to reverse the anonymization. Report any accidental identification to the maintainers so the file can be corrected.
• Withdrawal. Participants may withdraw. If a participant withdraws, the affected files are removed from the repository and the next version number is published. Please delete withdrawn files when notified.

Access and data use

License. MPSC-RespView is released under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0). You may share and adapt the data for non-commercial purposes, with attribution.

Access. The dataset is openly downloadable. By downloading or using it you also agree to the Terms of Use, in short:

1. Non-commercial use with attribution, as set out in the license.
2. No attempts to identify participants. Report any accidental identification to the maintainers.
3. Delete affected files when you are notified that a participant has withdrawn.
4. Cite the dataset and paper in any resulting publication.

The full terms are in DATA_USE_AGREEMENT.md and LICENSE.

Intended use and limitations

Intended for: training and evaluating respiratory-rate and respiratory-waveform estimators from video, including robustness to viewpoint, distance, breath holds and forced breathing.

Not intended for: identifying people, face analysis, demographic inference, or any clinical or diagnostic decision-making.

Limitations to keep in mind

• Faces are anonymized, so methods that rely on facial skin signals (for example face-based remote photoplethysmography) are not applicable. This dataset targets torso motion.
• Participants are healthy adults recorded at one site, and the dataset is small (73 recordings). Report results with subject-wise splits and confidence intervals.
• Outdoor recordings are single-camera one-minute clips from 5 subjects; indoor recordings are longer. Respiration-belt reference signals are not distributed with this release.
• Demographic attributes are not released, to protect participants.

Citation

Please cite the WACV 2027 paper whenever you use MPSC-RespView.

@inproceedings{sakib2027pixels,
  title     = {Pixels-to-Phase: Reliability-Aware Latent State-Space
               Modeling for Contactless Respiratory Estimation},
  author    = {Sakib, Shadman and Roy, Nirmalya},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on
               Applications of Computer Vision (WACV)},
  year      = {2027}
}

To cite the dataset itself (name and version) in addition to the paper:

@misc{sakib2026mpscrespview,
  title        = {{MPSC-RespView}: An RGB Video Dataset for Contactless
                  Respiratory Monitoring (v1.0)},
  author       = {Sakib, Shadman and Roy, Nirmalya},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/shadman-sakib/MPSC-RespView}}
}

Changelog

Version	Notes
v1.0	First release: 73 recordings (50 indoor, 23 outdoor).

Contact

Shadman Sakib	Ph.D. student, Department of Information Systems, UMBC · ssakib1@umbc.edu
Nirmalya Roy	Principal investigator, MPSC Lab, UMBC · nroy@umbc.edu

For access questions, corrections and takedown or withdrawal requests, email the maintainers above or open a Discussion on this repository.

MPSC Lab  ·  CYPRESS  ·  UMBC  ·  Paper website

_{Mobile, Pervasive and Sensor Computing (MPSC) Lab, University of Maryland, Baltimore County}

Acknowledgments

Data were collected at the Center for Research in Use-Inspired Cyber-Physical Systems (CYPRESS) and the Center for Real-time Distributed Sensing and Autonomy (CARDS), UMBC. We thank all participants.