LLVIP(IR dataset, ICCV 2021)

 

 

Figure 1: Samples of the LLVIP.

Datasets

• TNO Image Fusion Dataset: 261 pairs of images, few objects
• INO Videos Analytics Dataset & OTCBVS OSU Color-Thermal Database: few pedestrian
• CVC-14: images are already clear--no need for the ir images!!
• KAIST Multispectral Dataset: 
• FLIR Thermal Dataset

Figure 2: TNO and OSU dataset examples.

 

LLVIP Dataset

Figure 3: The establishment of the LLVIP dataset.

 

Image capture: captured with HIKVISION DS-2TD8166BJZFY-75H2F/V2 camera

- 15,488 pairs of visible-ir images from 26 diffrent locations

Figure 4: Image capture and registration.

 

Registration: 

ir camera and visible light camera sensors have sensors with different fields -- 

some registrations are applied for the two types of imagese have the same fieldd of vision and image size

 

fig. 5 shows the images before and after the registration. 

 

Annotations

since it is difficult to label very dark images, annotations are made on the ir version of the images and then copied to the vl images 

 

Advantages

• vl and ir images are synchronized
• the dataset is under low-light condition -- suitable for the image sfusion
• dataset contain a large number of annotations
• the quality of the images is very high

Disadvantages

• Most of the images are collected from a medium distance-- not suitable for the study of long or short distance detection

 

Discussion

 

 

---

References

[1] Jia, Xinyu, et al. "LLVIP: A visible-infrared paired dataset for low-light vision." Proceedings of the IEEE/CVF international conference on computer vision. 2021.

Footnotes