Skip to content

Latest commit

 

History

History
29 lines (24 loc) · 1.88 KB

README.md

File metadata and controls

29 lines (24 loc) · 1.88 KB
Raw

Coaxials dataset

  • Coaxials dataset was created in a research on the semantic segmentation task of automonous mobility robots operating near humans.
  • The dataset contains coaxial color (RGB) images and long wavelength infrared (LWIR) images captured with the same optical axis.
  • As a notable feature of the dataset, one of the annotated labels is "glass", such as windows, walls, and automatic doors. Glass is transparent in color images but opaque in LWIR images.
  • According to our paper, you can detect transparent glass using coaxial RGB and LWIR images.

Images and annotations

  • The images was captured with a coaxial camera, FIRplus, manufactured by the ViewPlus Corporation, Japan.

  • The dataset contains 17,023 image pairs, which is divided into 8,050 pairs of learning data, 8,373 pairs of evaluation data, and 600 pairs of verification data.

  • Nine labels were annotated: glass, pedestrian, road, aisle, tree, bicycle, car, devaluation, and curb.

  • The images were captured in eight lighting condition: daytime (clear sky, cloudy, rainy, snowy, indoor), evening, and night.

  • Examples of image# Images and annotations

Image example

Download dataset

  • Coaxials dataset is available on google drive. The file size is about 3 GB.

Citation

  • If you use Coaxials dataset in your work, please cite the following paper. 
    @PROCEEDINGS{coaxials-2019,
  title = {Simultaneous transparent and non-transparent object segmentation with multispectral scenes},
  author = {Okazawa, Atsuro and Takahata, Tomoyuki and Harada, Tatsuya},
  booktitle = {2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS2019)},
  year = {2019}
}