• Added 1708 frames of clearly labeled thermal images. (under driving_data)
• Labeling format in PASCAL VOC. (under driving_annotation)
• Covered a highway and a local driving scenario.
• Covered four object classes including vehicles, pedestrians, small animals(dogs), and bicycles.

This expansion pack is prepared specifically for training a YOU-ONLY-LOOK-ONCE(YOLO) network. All frames are labeled in the YOLO format. If you want to use this expansion pack for other purposes, images are still available for download but requires manual labeling.

The FREE FLIR Thermal Dataset for Algorithm Training can be downloaded here.

The YOLO format has the shape of
0 0.477734 0.545833 0.077344 0.175000

• The first integer 0 represents the object class.
• The float numbers represent x, y, width, and height.

A detailed description of the YOLO format can be found here.

FLIR's dataset includes 240 Dog bounding boxes. Training a learning algorithm using FLIR's dataset usually yields poor mAP on the dog class.
The following are the results from training a YOU-ONLY-LOOK-ONCE(YOLOv4) network using FLIR's dataset.

mAP@IoU=50 for Dog is 0.00%.

mAP@IoU=75 for Dog is 0.00%.

The generated bounding boxes are

The dogs are either missed or misclassified.

An example of the images in this expansion pack

An example of the new added dog bounding box (For demonstration purposes, box for other classes are not shown in the image, but they will appear in the annotation)

An example of the annotation
3 0.476953 0.543750 0.077344 0.170833

• Added 3,054 more annotated frames contain Dog captured by a FLIR A65 IR Temperature Sensor.
• Including long-fur and short-fur dogs with unique heat signature.
• All frames are of 640*512 and in .jpeg format making it consistent with the original FLIR dataset.
• All frames are 24-bit color images.

Images are taken from a recording session, so similar frames do exist. Omit them at your own demand.