LiFLabUC/NoseNet-training-data

Thermal face visibility and nose detection datasets

This repository contains the YOLO-format annotated thermal frames used to fine-tune the two YOLOv5m models described in the manuscript Open-standard infrared thermography implementation based on machine learning object identification for studying facial thermal variations at less restrictive settings.

Direct participant identifiers have been removed from the uploaded filenames. The private correspondence between original and pseudonymized filenames is retained by the data controller and is not part of this repository.

Dataset origin and annotation

The annotated material was generated by randomly sampling frames from a broader corpus of approximately 1,100 thermal recordings collected across three prior studies conducted with the same camera modules. The recordings lasted between one and five minutes. Sampling was designed to capture variation across participants, recordings, and experimental conditions before manual annotation.

Five laboratory labelers annotated the selected frames with LabelImg using shared written criteria and YOLOv5 bounding-box format. Manual annotation yielded 10,705 frames for the face-visibility task. Of the frames labeled as containing a face, 8,470 additionally received a nose bounding box and, after cropping to the facial region, constituted the nose-detection dataset.

The source recordings were obtained with FLIR Lepton 3.5 modules integrated into PureThermal Mini Pro JST-SR boards. The cameras produced native radiometric streams of 160 × 120 matrices containing 16-bit temperature values in hundredths of a kelvin. The PNG files in this repository are the exported or preprocessed images supplied to the YOLOv5 training workflow.

Dataset structure

face_visibility/
  data.yaml
  images/{train,val,test}/
  labels/{train,val,test}/
nose_detection/
  data.yaml
  images/{train,val,test}/
  labels/{train,val,test}/

Every .txt annotation uses normalized YOLO coordinates:

class_id x_center y_center width height

Splits and classes

Dataset	Classes	Train	Validation	Test	Total
Face visibility	0 = face, 1 = no_face	8,513	1,121	1,071	10,705
Nose detection	0 = nose	6,733	889	848	8,470

Face-visibility class distribution

A direct audit of the published label files produced the following image-level distribution:

Split	Total images	Only face	Only no_face	Both classes	Images containing no_face	Percentage containing no_face
Train	8,513	7,165	1,347	1	1,348	15.83%
Validation	1,121	944	177	0	177	15.79%
Test	1,071	904	167	0	167	15.59%
Total	10,705	9,013	1,691	1	1,692	15.81%

Thus, 1,691 of 10,705 images (15.80%) are annotated exclusively as no_face. When the single image containing annotations from both classes is included, 1,692 images (15.81%) contain at least one no_face annotation. The difference between these percentages reflects that single mixed-class image.

Each annotated frame occurs in only one subset. The supplied train, validation, and test directories therefore reproduce the frame-level partitions used for model development, with no repeated image between subsets. The detection objective is spatial: identifying usable facial information and locating the nose, rather than predicting participant identity or experimental condition.

Validation performed before publication

• Every image has a label file with the same stem.
• No orphan label files were found.
• All bounding-box coordinates are within the normalized range [0, 1].
• No unreadable PNG files were detected.
• The original train, validation, and test assignments were preserved.
• Published filenames follow a pseudonymized sequential naming scheme.

Recommended use

The datasets support reproduction of the reported YOLOv5 transfer-learning workflow and adaptation of pretrained object-detection models to thermal-imaging research. Researchers applying the workflow to a new camera, geometry, population, or recording environment should independently assess performance and fine-tune the models with representative frames when necessary.

Sensitive-data notice

These files contain low-resolution thermal images of research participants. They must be handled according to informed-consent terms, ethics approval, institutional requirements, and applicable data-use agreements. Do not use the dataset for identity recognition, surveillance, clinical diagnosis, or attempts to re-identify participants.

Related resources

• Model checkpoints
• Participant-independent validation dataset and results
• Project source code

License and access

No reuse license is granted by this card. Access, redistribution, and secondary use remain subject to the terms established by the authors, the responsible institutions, and the applicable participant authorizations.