FREE FLIR Thermal Dataset for Algorithm Training
The FLIR starter thermal dataset enables developers to start training convolutional neural networks (CNN), empowering the automotive community to create the next generation of safer and more efficient ADAS and driverless vehicle systems using cost-effective thermal cameras from FLIR.
Why Use FLIR Thermal Sensing for ADAS?
The ability to sense thermal infrared radiation, or heat, within the ADAS context provides both complementary and distinct advantages to existing sensor technologies such as visible cameras, Lidar and radar systems:
- With over 15 years of experience in automotive, FLIR has the only automotive-qualified thermal sensor that is deployed in over 500,000 cars today for driver warning systems.
- The FLIR thermal sensors can detect and classify pedestrians, bicyclists, animals and vehicles in challenging conditions including total darkness, fog, smoke, inclement weather and glare, providing a supplemental dataset beyond LiDAR, radar and visible cameras. The detection range is four times farther than typical headlights.
- When combined with visible light data and distance scanning data from LiDAR and radar, thermal data paired with machine learning creates a more comprehensive detection and classification system.
Updated FREE Thermal Dataset Coming Soon
|Content||Synced annotated thermal imagery and non-annotated RGB imagery for reference. Camera centerlines approximately 2 inches apart and collimated to minimize parallax|
|Images||>14K total images with >10K from short video segments and random image samples, plus >4K BONUS images from a 140 second video|
|Image Capture Refresh Rate||Recorded at 30Hz. Dataset sequences sampled at 2 frames/sec or 1 frame/ second. Video annotations were performed at 30 frames/sec recording.|
|Frame Annotation Label Totals||10,228 total frames and 9,214 frames with bounding boxes.
1. Person (28,151)
2. Car (46,692)
3. Bicycle (4,457)
4. Dog (240)
5. Other Vehicle (2,228)
|Video Annotation Label Totals||4,224 total frames and 4,183 frames with bounding boxes.
1. Person (21,965)
2. Car (14,013)
3. Bicycle (1,205)
4. Dog (0)
5. Other Vehicle (540)
|Driving Conditions||Day (60%) and night (40%) driving on Santa Barbara, CA area streets and highways during November to May with clear to overcast weather.|
|Capture Camera Specifications||IR Tau2 640x512, 13mm f/1.0 (HFOV 45°, VFOV 37°) FLIR BlackFly (BFS-U3-51S5C-C) 1280x1024, Computar 4-8mm f/1.4-16 megapixel lens (FOV set to match Tau2)|
|Dataset File Format||1. Thermal - 14-bit TIFF (no AGC)
2. Thermal 8-bit JPEG (AGC applied) w/o bounding boxes embedded in images
3. Thermal 8-bit JPEG (AGC applied) with bounding boxes embedded in images for viewing purposes
4. RGB - 8-bit JPEG
5. Annotations: JSON (MSCOCO format)
|Sample Results||mAP score of 0.587 (50% IoU) was obtained by fine tuning RefineDetect512 with this dataset and testing using holdout validation set. Details further explained in readme.|
|FLIR ADK Training and Development Settings||Use the FLIR ADK with default settings to begin data collection|