The DMR-IR (Database for Mastology Research - Infrared) is the most widely used public breast thermography dataset in research, utilized in 77% of studies in recent surveys. It represents the gold standard for breast cancer thermography research and has been instrumental in advancing computer-aided diagnosis systems.

The dataset was created by the Visual Lab at the Federal University Fluminense in collaboration with Hospital Universitario Antonio Pedro (HUAP). It has been approved by the Research Ethics Committee and registered with the Brazilian Ministry of Health under number CAAE: 01042812.0.0000.5243.

The dataset uses Dynamic Image Thermography (DIT), which has been shown to better differentiate between normal and abnormal breast tissue compared to static protocols. The protocol involves:

• Infrared Images: 20 sequential thermal images per patient captured during dynamic protocol
• Temperature Matrices: Numerical temperature data for each pixel in Celsius
• Digitized Mammograms: Corresponding mammography images for comparison
• ROI Masks: Pre-segmented breast regions for automated analysis
• Clinical Data: Age, race, personal history, family history, medical history
• Diagnosis Validation: 117 of 293 cases validated by biopsy (40%), with 74% of cancer cases confirmed

The DMR-IR dataset has been instrumental in advancing breast cancer detection through thermography:

• Camera Model: FLIR SC-620 infrared thermal camera
• Resolution: 640 × 480 pixels (meets American Standard for thermography)
• Thermal Sensitivity: Meets 50 mK NETD requirement
• Temperature Precision: 0.05°C (American Standard compliant)
• Image Formats: JPEG (visual) and thermal matrix format (numerical)
• Color Representations: Grayscale and colorized versions available
• Capture Environment: Temperature-controlled clinical setting

Access Link:

• DMR-IR Database Portal (Official)
• Kaggle Mirror Version

Usage Requirements:

• Free access for research purposes
• Proper citation required in publications
• Ethics committee approval obtained for dataset creation
• Patient consent forms signed

• Geographic Diversity: All patients from single institution in Brazil
• Validation Method: 60% of cases diagnosed by mammography alone, 40% confirmed by biopsy
• Temporal Scope: Dataset collected in single time period
• Model Generalization: High reported accuracies should be validated on external datasets
• Data Leakage Risk: Researchers should use patient-level splits, not image-level splits
• Protocol Standardization: Dynamic protocol not yet standard in clinical practice

This recent breast thermography dataset represents a valuable contribution from Colombia, providing high-quality thermal images with pathologically validated diagnoses. The dataset follows the American Academy of Thermology (AAT) protocol for standardized image capture.

Collected between 2021-2022 at San Juan de Dios Hospital in Cali, Colombia, this dataset addresses the critical scarcity of publicly available breast thermographic images. It has been used in state-of-the-art deep learning studies achieving over 97% accuracy in breast cancer classification.

The dataset follows the rigorous American Academy of Thermology protocol for breast thermography:

• Model: FLIR A300 (30 Hz)
• Lens: 18 mm with 25° Field of View
• Resolution: 320 × 240 pixels (16-bit)
• Thermal Sensitivity (NEdT): < 0.05°C @ +30°C (50 mK)
• Detector: Focal Plane Array (FPA), uncooled microbolometer
• IFoV: 1.36 mrad
• Field of View: Horizontal 25° × Vertical 18.8°
• Emissivity: 0.98 for skin
• Wavelength Range: Approximately 7.5-13 µm
• Working Distance: 1 meter from patient

All diagnoses in this dataset are validated through pathology studies following biopsy procedures, providing high-confidence ground truth labels:

• Malignant Pathology (PM): 35 patients with carcinoma at various stages
• Benign Pathology (PB): 84 patients with findings including:
  • Collagenized stroma
  • Fibroadenoma
  • Cysts
  • Adenosis
  • Apocrine metaplasia
  • Stromal fibrosis
  • Epithelial hyperplasia
  • Microcalcifications
  • Nodules

The comprehensive pathology classification makes this dataset particularly valuable for training models to distinguish between different types of breast abnormalities, not just binary normal/abnormal classification.

Recent studies using this dataset have demonstrated:

• Recent Collection: 2021-2022 capture period with modern equipment
• Multi-Position Views: Three angles per patient for comprehensive analysis
• Pathology Validated: All diagnoses confirmed through biopsy, not just mammography
• Detailed Classification: Specific benign pathology types documented
• Standardized Protocol: Follows internationally recognized AAT guidelines
• Controlled Environment: Detailed documentation of acquisition conditions
• Open Access: Freely available through Mendeley Data repository
• Good Balance: Includes both benign (70.6%) and malignant (29.4%) cases

Dataset Access:

• Mendeley Data Repository (Version 3)
• Dataset Description Paper (Data in Brief, 2024)

Supporting Institutions:

• Universidad Nacional Abierta y a Distancia (UNAD)
• Hospital San Juan de Dios de Cali
• Universidad del Valle
• Perception and Intelligent Systems Laboratory (PSI)

• Dataset Size: Smaller than DMR-IR (119 vs 293 patients), may benefit from augmentation
• Camera Resolution: 320×240 pixels lower than American Standard recommendation
• Geographic Diversity: Single institution in Colombia, may not represent global populations
• Static Protocol: Does not include dynamic cooling protocol like DMR-IR
• Demographic Details: Limited demographic breakdown beyond age range

Recommended Use Cases:

• Transfer learning from models trained on DMR-IR
• Multi-view analysis research
• Combined dataset studies with DMR-IR for increased diversity
• Validation of models trained on other datasets
• Benign vs. malignant classification research

OpenThermalPose is a pioneering open-source thermal human pose estimation dataset that addresses critical challenges in computer vision, particularly insufficient illumination and privacy concerns that plague visible-domain pose estimation systems.

The dataset was created by the Intelligent Systems and Signal Processing (IS2AI) Laboratory at Nazarbayev University, Kazakhstan. It provides comprehensive annotations following the MS COCO keypoint format, making it immediately compatible with existing pose estimation frameworks and models.

Each human instance is annotated with 17 anatomical keypoints following the MS COCO format:

This standardized format ensures compatibility with popular frameworks like YOLOv8-pose, YOLO11-pose, and other COCO-based pose estimation models.

The dataset captures diverse scenarios to ensure model robustness:

• Fitness Exercises: Various workout poses and movements
• Multi-Person Scenes: Multiple humans interacting in single frames
• Outdoor Walking: Natural gait and movement patterns
• Different Locations: Indoor and outdoor environments
• Weather Conditions: Various ambient temperature scenarios
• Lighting Conditions: From complete darkness to daylight
• Occlusion Scenarios: Partial body occlusions and challenging poses

The authors provide pre-trained YOLOv8-pose models as baselines:

All models are trained from scratch on the thermal dataset and are available for download with full training configurations.

The dataset enables various privacy-preserving and low-light applications:

Dataset Structure:

• Train/Val/Test Split: Standard splits provided
• Image Format: PNG (lossless thermal image storage)
• Label Format: TXT files (YOLO format) and JSON (COCO format)
• YAML Configuration: Ready-to-use training configuration files

Easy Integration:

GitHub Repository:

• Official GitHub Repository
• IEEE FG 2024 Paper

What's Included:

• Complete annotated dataset (train/val/test)
• Pre-trained YOLOv8-pose models (all sizes)
• Training scripts and configuration files
• Evaluation metrics and benchmarks
• Documentation and usage examples
• Data loading utilities

Key Advantages:

• Privacy-Preserving: Thermal imaging obscures facial features and identity
• Illumination-Invariant: Works equally well in darkness and bright light
• Open Source: Free access to data, code, and models
• Standard Format: COCO compatibility enables wide framework support
• Diverse Scenarios: Multiple activities, locations, and conditions
• Ready-to-Use: Pre-trained models and complete pipeline provided

See Also: OpenThermalPose2

An extended version (OpenThermalPose2) is now available with 11,391 images of 170 subjects and 21,125 annotated instances—nearly double the original dataset size with even more diverse poses and scenarios.

OpenThermalPose2 is the extended version of the original OpenThermalPose dataset, featuring nearly double the data, subjects, and pose variations. It represents the most comprehensive open-source thermal human pose dataset available.

• More Fitness Exercises: Expanded variety of workout poses
• More Multi-Person Scenes: Complex interactions and group activities
• More Locations: Additional indoor and outdoor environments
• More Weather Conditions: Extended seasonal and temperature variations
• More Challenging Poses: Difficult occlusions and body configurations

Models trained on OpenThermalPose2 consistently outperform those trained on the original dataset:

• Better generalization to unseen scenarios
• Improved accuracy on challenging poses
• Enhanced multi-person tracking
• Reduced false positives in cluttered scenes

The dataset now includes both YOLOv8-pose and YOLO11-pose baseline models, with TensorRT optimized versions for real-time deployment on NVIDIA Jetson devices.

• GitHub Repository (includes v2)
• IEEE DataPort
• Hugging Face Datasets

This comprehensive dataset from PhysioNet provides thermal and visible facial images paired with oral temperature measurements from over 1,000 subjects. It was specifically designed for developing and validating automated temperature screening systems, particularly relevant for pandemic response, fever detection at entry points, and public health monitoring applications.

The dataset addresses the critical need for non-contact temperature screening methods by providing ground truth oral temperature measurements alongside facial thermal and visible images. This pairing enables researchers to develop algorithms that can accurately predict core body temperature from facial thermal signatures, a crucial capability for high-throughput screening scenarios.

Public Health Screening:

• Fever Detection: Identify individuals with elevated temperature at entry points
• Mass Screening: Non-contact temperature assessment in crowds
• Pandemic Response: First-line screening for infectious diseases
• Healthcare Facilities: Pre-screening patients and visitors

Infrastructure Deployment:

• Airports & Borders: International traveler screening
• Office Buildings: Employee health monitoring at entrances
• Schools & Universities: Student screening for illness
• Event Venues: Large gathering temperature checks

Research & Development:

• Algorithm Validation: Test accuracy of thermal screening systems
• ROI Optimization: Identify best facial regions for temperature estimation
• Multi-Modal Fusion: Combine thermal and visible data for improved accuracy
• Calibration Studies: Understand environmental and demographic factors

Optimal Facial ROIs for Temperature:

• Inner Canthi: Corners of eyes near nose - most reliable for core temperature
• Forehead: Large area, but affected by environmental factors
• Periorbital Region: Area around eyes with consistent thermal patterns
• Nose Tip: Alternative region, though more variable

Factors Affecting Accuracy:

• Ambient Temperature: Environmental conditions impact facial temperatures
• Recent Activity: Physical exertion elevates skin temperature
• Facial Accessories: Glasses, hats, scarves obstruct thermal imaging
• Distance & Angle: Camera positioning affects measurement accuracy
• Skin Emissivity: Varies with skin tone and moisture

Large Scale:

• Over 1,000 subjects provide statistical significance
• Diverse population for generalization testing
• Multiple conditions and variations captured

Ground Truth Reference:

• Oral temperature gold standard for validation
• Quantitative accuracy assessment possible
• Enables calibration of thermal systems

Multi-Modal Data:

• Thermal and visible images for fusion approaches
• Facial features help with region detection
• Enables demographic correlation studies

PhysioNet Platform:

• Trusted medical research repository
• Standardized access and documentation
• Community of medical researchers
• Version control and reproducibility

Algorithm Development:

• Develop automated ROI detection for optimal temperature regions
• Create regression models for core temperature prediction
• Build classification systems for fever/no-fever detection
• Implement compensation for environmental factors

Multi-Modal Learning:

• Fuse thermal and visible data for improved accuracy
• Use facial landmarks from RGB for thermal ROI guidance
• Develop attention mechanisms focusing on relevant facial regions
• Cross-modal feature learning and domain adaptation

Clinical Studies:

• Validate thermal screening sensitivity and specificity
• Study demographic variations in thermal patterns
• Analyze effect of ambient conditions on accuracy
• Develop clinical guidelines for thermal screening

Camera Requirements:

• Thermal sensitivity: ≤0.05°C recommended for medical screening
• Resolution: Higher resolution for better ROI detection
• Calibration: Regular calibration against blackbody reference
• Field of view: Appropriate for facial capture at screening distance

Processing Pipeline:

• Face Detection: Locate face in thermal or visible image
• ROI Extraction: Identify optimal temperature measurement regions
• Temperature Calculation: Extract max/mean temperature from ROI
• Compensation: Adjust for ambient conditions if needed
• Classification/Regression: Predict core temperature or fever status

Dataset Access:

• PhysioNet Repository - Official Access

PhysioNet Registration:

• Free registration required for PhysioNet access
• Complete credentialing for some restricted datasets
• Agree to data use agreements
• Cite properly in publications

What's Included:

• Thermal facial images in standard format
• Visible light photographs
• Oral temperature measurements
• Subject demographic information
• Capture metadata (time, conditions)
• Documentation and data dictionary

Known Limitations:

• Environmental Sensitivity: Facial temperatures affected by ambient conditions
• Individual Variation: Baseline temperatures vary between people
• Accessory Interference: Glasses, masks can obstruct thermal readings
• Distance Dependency: Accuracy decreases with distance from camera
• Not Diagnostic: Screening tool only, not for medical diagnosis

Best Practices for Use:

• Use as first-line screening, confirm with clinical thermometer
• Control environment when possible (avoid drafts, direct sun)
• Allow subjects to acclimate to environment before measurement
• Set appropriate thresholds based on validation studies
• Regular calibration and system validation
• Train operators on proper use and limitations

• Deep Learning ROI Detection: Automated optimal region identification
• Environmental Compensation: ML models to adjust for ambient factors
• Individual Baseline Learning: Personalized temperature profiles
• Multi-Camera Fusion: Combine multiple views for robustness
• Temporal Analysis: Track temperature changes over time
• Integration with Symptoms: Combine thermal with other health indicators

LWIRPOSE is a novel dataset that bridges the gap between RGB and Long-Wave Infrared (LWIR) thermal imaging for 2D human pose estimation. It provides nearly aligned RGB-Thermal image pairs with comprehensive 2D human pose annotations, enabling research in cross-modal learning and multi-spectral pose estimation.

The dataset addresses the critical challenge of limited annotated thermal pose data by providing paired visible and thermal images. This pairing allows researchers to leverage knowledge from well-established RGB pose estimation models and transfer it to the thermal domain.

Pairing Strategy:

• Nearly Aligned: RGB and LWIR cameras positioned to capture same scene from similar viewpoints
• Temporal Synchronization: Images captured at same time instances
• Spatial Correspondence: Careful calibration ensures pixel-level correspondence
• Shared Annotations: Same 2D pose annotations applicable to both modalities

Pose Annotation Details:

• 2D anatomical keypoints for human joints
• Bounding boxes for person detection
• Occlusion and visibility flags
• Quality control through manual verification

• Paired Data: Enables supervised cross-modal learning approaches
• Benchmark for Transfer Learning: Test RGB-to-thermal and thermal-to-RGB transfer
• Complementary Information: RGB captures texture/color, thermal captures heat signatures
• Real-World Scenarios: Demonstrates practical multi-modal sensing systems
• Smaller Scale: Manageable size for rapid prototyping and experiments
• Focus on Pose: Specifically designed for pose estimation, not general detection

Camera Setup:

• Dual-camera rig with RGB and LWIR sensors
• Calibrated for geometric alignment
• Synchronized capture to minimize temporal differences
• Fixed relative positions for consistent pairing

Alignment Challenges:

• Transfer Learning Research: Study how RGB models transfer to thermal domain
• Multi-Modal Fusion: Combine RGB and thermal features for robust pose estimation
• Domain Gap Analysis: Quantify differences between RGB and thermal pose estimation
• Sensor Selection Studies: Evaluate when to use RGB vs. thermal vs. both
• Adaptation Techniques: Test unsupervised and semi-supervised adaptation methods
• Baseline Comparisons: Establish RGB-thermal benchmarks for future work

Dataset Access:

• GitHub Repository - Official
• arXiv Paper - Technical Details

What's Included:

• RGB image files
• LWIR thermal image files
• 2D pose annotations in standard format
• Camera calibration parameters
• Train/validation/test splits
• Documentation and usage examples

Current Limitations:

• Scale: Smaller than pure RGB or thermal-only datasets
• Diversity: Limited number of subjects and scenarios
• Alignment: Not perfectly aligned at pixel level
• Indoor Bias: May have limited outdoor/varying weather conditions

Future Research Opportunities:

• Expand dataset with more subjects and diverse scenarios
• Improve alignment through better calibration techniques
• Add temporal sequences for video-based pose estimation
• Include depth information for 3D pose estimation
• Capture in varying environmental conditions

IPHPDT (Identity-Preserved Human Posture Detection in Thermal Images) is a large-scale infrared thermal dataset specifically designed for privacy-preserving human posture recognition. With approximately 75,000 images, it represents one of the largest publicly documented thermal datasets focused on activity recognition and posture classification.

The dataset addresses critical privacy concerns in surveillance and monitoring systems by using thermal imaging, which naturally obscures facial features and personal identifying characteristics while maintaining the ability to recognize human postures and activities. This makes it particularly valuable for healthcare, elderly care, and smart home applications where privacy is paramount.

Why Thermal Imaging for Privacy:

• No Facial Features: Thermal cameras detect heat signatures, not visual details - faces are not recognizable
• No Clothing Details: Personal clothing, accessories, and identifying marks are not visible
• No Skin Tone: Racial and ethnic characteristics are not captured
• Universal Heat Signature: All humans appear similar thermally regardless of appearance
• Darkness Operation: Works in complete darkness, allowing monitoring without intrusive lighting
• GDPR Compliant: Inherently privacy-preserving, reducing regulatory concerns

Healthcare & Medical:

• Fall Detection: Real-time detection of falls in elderly care facilities
• Patient Monitoring: Track posture changes in hospital beds
• Rehabilitation: Monitor patient exercises and posture compliance
• Sleep Studies: Analyze sleep postures without disturbing patients

Smart Home & Assisted Living:

• Activity Recognition: Understand daily living activities
• Emergency Detection: Detect unusual postures indicating distress
• Energy Management: Occupancy detection for HVAC and lighting
• Security: Intrusion detection without privacy concerns

Research Applications:

• Activity Recognition Algorithms: Train and test classification models
• Privacy-Preserving AI: Develop ethical monitoring systems
• Temporal Analysis: Study posture transitions and movement patterns
• Robustness Testing: Evaluate models under various conditions

Capture Protocol:

• Multiple camera viewpoints to capture postures from different angles
• Controlled indoor environment with consistent temperature
• Multiple subjects performing each posture category
• Natural posture variations to increase diversity
• Different distances from camera to test scale invariance

Quality Control:

• Manual verification of posture labels
• Removal of ambiguous or transitional poses
• Balanced representation across all four categories
• Multiple examples per subject-posture combination

Suitable Architectures:

• CNNs: ResNet, VGG, MobileNet for posture classification
• Vision Transformers: ViT for global posture understanding
• Temporal Models: LSTMs, GRUs for posture sequence analysis
• Lightweight Models: EfficientNet, MobileNet for edge deployment

Training Strategies:

• Transfer learning from RGB posture datasets
• Data augmentation (rotation, scaling, thermal noise)
• Multi-view fusion for improved accuracy
• Temporal consistency constraints for video sequences

According to the paper, various deep learning models have been tested on IPHPDT with strong classification performance:

• High accuracy across all four posture categories
• Robust performance across different subjects
• Good generalization to unseen viewpoints
• Real-time inference capability demonstrated

Publication:

• MDPI Sensors Journal Paper (2023)

Dataset Access:

The dataset is described in detail in the MDPI Sensors paper. For access to the actual dataset files, researchers should contact the authors directly through the paper or check for supplementary data availability on the journal website.

Dataset Limitations:

• Indoor Only: Captured in controlled indoor environments
• Limited Postures: Four categories may not cover all real-world postures
• Transitional Poses: May lack images of posture transitions
• Access Restrictions: May require author contact for full dataset
• Single Activity: Each image shows single posture, not complex activities

Practical Considerations:

• Models may need adaptation for outdoor or varying temperature environments
• Multiple postures in single frame not explicitly covered
• Clothing thickness may affect thermal signatures
• Camera placement and angle critical for deployment

• Expand Posture Classes: Add more complex activities and transitions
• Multi-Person Scenarios: Include multiple people in single frame
• Temporal Extensions: Add video sequences for action recognition
• Environmental Diversity: Capture in various indoor and outdoor settings
• Edge Deployment: Optimize models for embedded thermal cameras
• Multi-Modal Fusion: Combine with other privacy-preserving sensors

The POP (Partially Occluded Person) Thermal Dataset is a specialized UAV-based infrared thermal imaging dataset designed for detecting people in outdoor environments where partial occlusions are common. Captured from aerial platforms, this dataset addresses the unique challenges of person detection from drone perspectives in scenarios where individuals may be partially hidden by vegetation, structures, or terrain features.

Published in Nature Scientific Data (2025), this dataset fills a critical gap in UAV-based person detection research. Unlike ground-level surveillance, aerial thermal imaging from drones presents unique challenges including varying altitudes, bird's-eye view perspectives, and environmental occlusions that are distinctly different from traditional surveillance scenarios.

Search & Rescue Operations:

• Missing Person Location: Detect people lost in wilderness or disaster areas
• Disaster Response: Find survivors in collapsed buildings or debris
• Night Operations: Thermal imaging enables 24/7 search capabilities
• Large Area Coverage: UAVs can rapidly scan extensive search zones

Security & Surveillance:

• Border Patrol: Detect unauthorized crossings in remote areas
• Perimeter Security: Monitor large facilities or restricted zones
• Event Surveillance: Aerial monitoring of crowds and gatherings
• Critical Infrastructure: Protect power plants, pipelines, military bases

Research & Development:

• Occlusion-Robust Detection: Train models to handle partial visibility
• Multi-Scale Detection: Handle persons at various distances from camera
• UAV Vision Systems: Develop autonomous drone surveillance
• Thermal Image Processing: Advance aerial thermal analysis techniques

Occlusion Types Covered:

• Vegetation: Trees, bushes, tall grass obscuring parts of people
• Structures: Buildings, walls, vehicles blocking thermal signatures
• Terrain: Hills, rocks, uneven ground causing partial visibility
• Environmental: Shadows, mixed thermal backgrounds

Capture Conditions:

• Various times of day (different ambient temperatures)
• Different weather conditions
• Multiple UAV flight patterns and altitudes
• Diverse outdoor environments (urban, rural, natural)
• Various levels of occlusion severity

The dataset includes detailed annotations for training and evaluating person detection models:

• Bounding Boxes: Coordinates for each detected person (even partially occluded)
• Occlusion Labels: Indicators of occlusion severity and type
• Person Count: Number of people in each frame
• Visibility Percentage: Estimated percentage of body visible
• Quality Flags: Image quality and detection difficulty ratings

UAV & Camera Setup:

• Thermal infrared camera mounted on commercial or research drones
• Long-wave infrared (LWIR) spectrum detection
• High-resolution thermal sensors for detailed heat signatures
• Stabilized gimbal mounting for image stability
• GPS-tagged images for location tracking

Image Properties:

• Thermal infrared spectrum (7-14 μm typical)
• Variable resolution depending on altitude
• 16-bit thermal data or processed 8-bit images
• Metadata including altitude, GPS, timestamp

Detection Architectures:

• YOLO Family: Real-time detection for onboard UAV processing
• Faster R-CNN: High accuracy for post-processing analysis
• Feature Pyramid Networks: Multi-scale detection for varying altitudes
• Attention Mechanisms: Focus on partially visible thermal signatures

Training Strategies:

• Data augmentation: rotation, scaling, thermal noise injection
• Transfer learning from ground-based thermal person detection
• Hard negative mining for thermal false positives (hot surfaces)
• Occlusion-aware loss functions
• Multi-task learning (detection + occlusion estimation)

Publication:

• Nature Scientific Data Paper (2025)

Dataset Access:

The dataset access information is provided in the Nature Scientific Data publication. Researchers should refer to the paper for download links, usage agreements, and any restrictions on dataset use.

Key Advantages:

• Unique Perspective: Aerial UAV viewpoint rarely available in other datasets
• Occlusion Focus: Specifically designed for partial visibility scenarios
• Real-World Relevance: Directly applicable to search & rescue operations
• Scale Diversity: Various altitudes provide multi-scale training data
• Outdoor Emphasis: Covers challenging outdoor thermal environments
• Recent Publication: Reflects current UAV and thermal camera technology

Limitations to Consider:

• Specialized Scenario: May not generalize to all person detection tasks
• Outdoor Only: Limited to outdoor environments
• UAV-Specific: Results may not transfer to ground-based or other platforms
• Annotation Complexity: Partial occlusions challenging to annotate consistently
• Environmental Dependency: Performance may vary with weather and terrain

• Temporal Sequences: Extend to video for tracking occluded people
• Multi-Spectral Fusion: Combine thermal with RGB for improved detection
• Autonomous Navigation: Use for UAV path planning around obstacles
• Activity Recognition: Classify actions of partially visible people
• 3D Reconstruction: Estimate 3D positions from multiple UAV viewpoints
• Edge Processing: Optimize models for onboard UAV computation

This Kaggle-hosted thermal imaging dataset provides a straightforward, accessible resource for training and testing thermal human detection models. Pre-formatted with YOLO-style bounding box annotations, it's designed for researchers and developers who want to quickly prototype thermal person detection systems without extensive preprocessing.

The dataset serves as an excellent starting point for learning thermal image processing and object detection, offering immediate compatibility with popular detection frameworks like YOLOv5, YOLOv8, YOLOv11, and other YOLO variants. The community-driven nature of Kaggle also means active discussions, shared notebooks, and collaborative improvements.

The dataset uses the standard YOLO annotation format, making it plug-and-play with YOLO detection frameworks:

Format Structure:

• One text file per image: Each .txt file corresponds to an image
• Format: class_id x_center y_center width height (all normalized 0-1)
• Multiple objects: One line per object/person in the image
• Normalized coordinates: Values relative to image dimensions

Ease of Use:

• No Preprocessing: Annotations ready to use immediately
• Standard Format: Compatible with all major YOLO implementations
• Fast Download: Kaggle's infrastructure ensures quick access
• Free Hosting: No storage costs or access fees
• API Integration: Programmatic download in automated pipelines

Community Support:

• Active Discussions: Forum for questions and solutions
• Shared Notebooks: Pre-built training examples and experiments
• Version Control: Dataset updates tracked and documented
• Reproducibility: Fixed versions ensure consistent experiments

Rapid Prototyping:

• Quick Iteration: Test ideas without annotation overhead
• Baseline Models: Establish performance benchmarks quickly
• Transfer Learning: Pre-train on this before fine-tuning on custom data
• Algorithm Testing: Compare detection approaches efficiently

The YOLO format makes this dataset compatible with numerous detection frameworks:

Educational:

• Learn thermal image processing and analysis
• Understand YOLO detection architecture
• Practice data preprocessing and augmentation
• Experiment with hyperparameter tuning

Research:

• Baseline for comparing new detection algorithms
• Pre-training before transfer learning
• Testing domain adaptation techniques
• Benchmarking inference speed optimizations

Development:

• Prototype thermal surveillance systems
• Develop edge device detection applications
• Create proof-of-concept demonstrations
• Test deployment pipelines

Access Methods:

• Web Download: Click download button on Kaggle dataset page
• Kaggle API: Command-line download for automation
• Kaggle Notebooks: Direct access in cloud environment
• Version Control: Pin specific versions for reproducibility

Kaggle Notebooks Features:

• Free GPU/TPU for training (limited hours per week)
• Pre-installed ML libraries and frameworks
• Easy sharing and collaboration
• Integrated with dataset (no download needed)

YOLOv8 Training (Ultralytics):

Dataset Access:

• Kaggle Dataset Page

Additional Resources:

• Community notebooks with training examples
• Discussion forum for Q&A
• Dataset metadata and descriptions
• Version history and updates

Getting Started:

• Create free Kaggle account
• Accept dataset terms
• Download or use in Kaggle Notebook
• Start training immediately

For Training:

• Data Augmentation: Add rotation, flip, brightness adjustments
• Validation Split: Ensure proper train/val separation
• Hyperparameter Tuning: Experiment with learning rate, batch size
• Early Stopping: Monitor validation metrics to prevent overfitting

For Deployment:

• Model Optimization: Quantization, pruning for edge devices
• Inference Testing: Benchmark on target hardware
• Error Analysis: Study false positives/negatives
• Domain Adaptation: Fine-tune on your specific thermal camera

Dataset Limitations:

• General Purpose: May not cover specialized scenarios
• Unknown Origin: Limited metadata about capture conditions
• Camera Specs: Source camera types may not be documented
• Annotation Quality: Community-contributed, may have inconsistencies

Recommended Use:

• Learning and educational purposes
• Initial prototyping and algorithm testing
• Transfer learning baseline before custom data
• Quick feasibility studies

Not Recommended For:

• Production systems without validation on target domain
• Safety-critical applications (medical, autonomous vehicles)
• Scenarios requiring specific thermal camera calibration
• Applications needing detailed metadata

For more comprehensive thermal person detection research, consider combining this dataset with:

• OpenThermalPose: Add pose estimation capabilities
• IPHPDT: Expand with posture classification
• POP Thermal: Add aerial/UAV perspectives
• Custom Collections: Capture your own data for specific use cases

Combining datasets helps create more robust models that generalize better across different thermal cameras, environments, and scenarios.