TFD68: A Fully Annotated Thermal Facial Dataset with 68 Landmarks, Pose Variations, Occlusions, Expressions, and Paired Visual Images

Overview
TFD68 is a controlled‑environment thermal facial dataset comprising 28,496 thermal images from 137 participants, each paired one‑to‑one with a corresponding visible‑light facial image captured under identical pose and occlusion conditions. Every thermal image includes 68 facial landmarks, per‑pixel thermal maps, and occlusion annotations. Head‑pose coverage spans 65 combinations (13 yaw × 5 pitch), and seven facial expressions (neutral, happiness, sadness, anger, fear, disgust, surprise) are included to support pose‑ and occlusion‑robust modelling. Images were acquired using a FLIR A400 (320×240, 7.5–14 μm) with paired visible images at 640×480 in a temperature‑controlled lab. (Source: TFD68 paper) [DOI:10.1145/3757376.3771410]

Data structure
Subject‑disjoint train/validation/test splits are provided at 60/20/20. Folders: “thermal/” (thermal images), “visual/” (visible images), “thermalfile/” (temperature/metadata). COCO‑format JSON annotations are supplied as train.json, val.json, test.json, and challenging.json, with seven categories (jawline, eyebrows, nose, right eye, left eye, mouth, face bbox). Landmark tuples follow (x, y, v), where v ∈ {1: occluded, 2: un‑occluded}. A separate “challenging” subset contains difficult cases (only images with face bbox were manually verified). (Source: README – Read Before Use)

Ethics
Ethical approval was granted by the Heriot‑Watt University Engineering & Physical Sciences Ethics Committee on 19 July 2023 (ref: 2023‑5418‑8035). All participants were 18+ and provided informed consent via the online consent form and information sheet.

Access & usage restrictions
Because the dataset contains identifiable facial images, files are released under Restricted Access for non‑commercial research use. Publication or public display of example images is limited to Subject IDs 5 and 137 (research team members) to reduce the risk of future takedown requests under GDPR (UK) and PDPA (Malaysia). Redistribution is prohibited. Users must comply with the “READ BEFORE USE” conditions. (Source: README – Read Before Use)

Custom Licence – Non‑Commercial Research Use Only
This dataset contains identifiable thermal and visible facial images and is therefore protected under GDPR (UK) and PDPA (Malaysia). Access is restricted to non‑commercial academic research only. Redistribution, reposting, or sharing the dataset, in whole or in part, is strictly prohibited.
Only images belonging to Participant IDs 5 and 137 (members of the research team) may be used in publications or public‑facing materials. Images of all other participants may be used for analysis only and must not be displayed or published.
The dataset must not be used for identity recognition, surveillance, commercial use, or re‑identification activities. Users must comply with the conditions set out in the accompanying “READ BEFORE USE” document. Any further distribution or use outside these conditions requires explicit written permission from Heriot‑Watt University Malaysia.

How to cite
Please cite the dataset publication:
Yean Chun Ng, F. C. M. Choong, Alexander G. Belyaev, Joon Huang Chuah, Shahrel A. Suandi, and Bhuvendhraa Rudrusamy. “TFD68: A Fully Annotated Thermal Facial Dataset with 68 Landmarks, Pose Variations, Per‑Pixel Thermal Maps, Visual Pairs, Occlusions, and Facial Expressions.” SIGGRAPH Asia Technical Communications, 2025. DOI: 10.1145/3757376.3771410.

Contact (data custodian): Dr Bhuvendhraa Rudrusamy – [email protected]