Datasets:
protocol dict | primary_operational_results dict | secondary_box_results dict | external_cnf_subset dict | annotation_note stringclasses 1
value |
|---|---|---|---|---|
{
"participants": 11,
"recordings": 133,
"frames": 220,
"positive_frames": 210,
"no_face_frames": 10,
"frames_per_participant": 20,
"study_1_participants": 8,
"external_cnf_participants": 3,
"exact_training_frame_overlap": 0,
"image_size": 640,
"operating_confidence": 0.5,
"nms_iou": 0.45,
"we... | {
"face_sensitivity": {
"successes": 210,
"total": 210,
"percent": 100
},
"no_face_specificity": {
"successes": 9,
"total": 10,
"percent": 90
},
"face_classification_accuracy": {
"successes": 219,
"total": 220,
"percent": 99.5455
},
"complete_nasal_box_retention_by_face... | {
"face_iou50_percent": 81.43,
"face_no_face_map50_percent": 75.9,
"isolated_nose_iou50_percent": 95.71,
"isolated_nose_ap50_percent": 95.82,
"end_to_end_nose_iou50_percent": 72.38,
"end_to_end_nose_ap50_percent": 60.77
} | {
"face_classification_percent": 100,
"isolated_nose_iou50_percent": 95,
"end_to_end_nose_iou50_percent": 90
} | Bounding boxes were drawn rapidly by one annotator; box metrics are secondary diagnostics. |
NoseNet independent validation dataset
This repository contains the 220 manually annotated thermal frames used for the post-hoc independent validation of the frozen NoseNet face-visibility and nose-localization checkpoints. It accompanies the manuscript Open-standard infrared thermography implementation based on machine learning object identification for studying facial thermal variations at less restrictive settings.
The scripts, validation protocol and aggregated results used for this evaluation are available in the NoseNetValidation GitHub repository.
Independence from model development
The validation dataset contains 220 frames sampled from 133 recordings belonging to 11 participants who did not contribute images to model training. Eight participants came from Study 1 recordings not represented in training, and three came from a separate study acquired with the same thermal-camera technology. Exact SHA-256 comparison found no validation image reused in the training set. The published checkpoints were frozen before manual annotation, and no retraining or threshold adjustment was performed after annotation.
Sample composition
| Item | Count |
|---|---|
| Participants | 11 |
| Recordings | 133 |
| Total frames | 220 |
| Frames with visible face and nose | 210 |
| Frames without a visible face | 10 |
| Frames per participant | 20 |
| Exact training-frame overlap | 0 |
Repository structure
images/ # 220 thermal PNG frames
labels/ # combined YOLO annotations
manifest.csv # pseudonymized grouping and provenance fields
results/per_frame_results.csv # de-identified frame-level predictions and metrics
results/*.json # aggregate and full de-identified summaries
docs/ # selection, evaluation and availability protocols
scripts/evaluate_external_validation.py
classes.txt
YOLO class identifiers are 0 = face, 1 = nose, and 2 = no_face. Two legacy positive-frame annotation files used duplicated class identifiers; for publication, their class identifiers were normalized deterministically by assigning the larger box to face and the smaller box to nose. Bounding-box coordinates were not changed. This operation is recorded in manifest.csv.
Participant and recording identifiers are pseudonyms. Original identifiers, source paths and the private correspondence table are not included.
Frozen-checkpoint evaluation
At an operating confidence of 0.50 and NMS IoU of 0.45, the face-visibility model detected all 210 positive frames and rejected 9 of 10 no-face frames, for an overall classification accuracy of 219/220 (99.55%). With manually annotated facial crops, the nose detector produced a detection in 207/210 frames (98.57%). In the full two-stage workflow, nose localization was produced in 209/210 positive frames (99.52%), and the predicted nose-box center fell within the manually annotated nose region in 209/210 frames.
Bounding boxes were drawn rapidly by one annotator. IoU and AP50 results are therefore included as secondary localization diagnostics, not as definitive estimates of bounding-box precision. See results/ and docs/ for the complete metrics and protocol.
Related resources
Ethical use and license
These files contain low-resolution thermal images of research participants. Use, redistribution and secondary analysis must comply with participant consent, ethics approval, institutional requirements and applicable data-use agreements. Do not use the data for identity recognition, surveillance, clinical diagnosis or attempts to re-identify participants. No reuse license is granted by this card.
- Downloads last month
- 61