Dataset Name

 

Resolution

 

Wavelength band

 

Bits per pixel

 

RGB images also

 

Number of images/frames

 

Number of Classes

 

Remarks

 

OSU-TP (Davis and Keck 2005) (01)

 

360 x 240

 

LWIR

 

8

 

N

 

284 images, 10 sequences

 

1 (Pedestrian)

 

Very less number of images, and low resolution.

 

OSU-CT (Davis and Sharma 2007) (03)

 

360 x 240

 

LWIR

 

8

 

Y

 

17089 images, 6 sequences

 

1 (Humans)

 

Results stored in CV markup language, so annotations availability needs to be checked

 

Terravic Motion (Miezianko) (05)

 

320 x 240

 

LWIR

 

8

 

N

 

10 small video collections

 

1 (Humans)

 

Annotations are not available.

 

LITIV (Torabi et al. 2012)

(Trimodal)

 

 

320 x 240

 

 

LWIR

 

 

8

 

 

Y

 

5724 annotated frames divided into 3 separate indoor scenes

 

 

1 (Humans)

 

Mainly  for registering infrared and visible people appearing at different planes, masks and annotations available.

 

ASL-TID (Portmann et al. 2014)

 

324 x 256

 

LWIR

 

8/16

 

N

 

4381 (with objects) + 2418 (background)

 

3 (Humans, cat and horse)

 

Can be used for very small object detection (check the examples shown in dataset page)

 

 

BU-TIV (Wu et al. (2014))

 

 

Upto 1024 x 1024

 

 

MWIR

 

 

16

 

 

N

 

 

11 different data sequences

 

People, Cars, Motorcycles, Bicycles and Bats

 

Will be helpful for single and  multi-object tracking, but the annotations and datafiles are not clear.

 

KAIST-RGBT (Hwang et al. 2015)

 

Dataset Demo Video

 

 

640 x 512 (320x x 256)

 

 

LWIR

 

 

8

 

 

Y

 

95k Colour & Thermal pairs,

with 103, 128 dense annotations, 1182 unique pedestrians

 

3

(person, people, cyclist)

 

 

This is the dataset we are using

 

 

LTIR (Berg et al. 2016 )

 

 

Upto 1920 x 480

 

 

LWIR

 

 

8/16

 

 

N

 

Average sequence length 563

 

Humans, Dog, Car, Horse, Quadcopter, Rhinoceros

 

 

Good number of images for human detection  (but in different resolutions and environments), but less for other cases.

 

Deep Thermal Imaging (Youngjun et al. 2018) (3)

Paper

 

 

75 x 75

 

 

LWIR

 

 

 

 

Y

 

15 indoor materials and 17 outdoor materials

 

14,860 indoor + 26,584 outdoor images

 

None of the image categories present in this dataset will be useful for our problem.