IRS Scene Understanding Dataset

IRS stands for Indoor Robotics Stereo, which is an open image dataset for indoor robot vision tasks. The images contain surface normals and disparity maps, a total of 103,316 images, and the image content is mainly four types of indoor scenes: home, office, store and restaurant.

The dataset includes:

• Indoor scenes: 31,145 home images, 43,417 office images, 22,058 restaurant images, 6,696 store images
• Target objects: desks, chairs, sofas, glass, mirrors, beds, bedside tables, lamps, wardrobes, etc.
• Brightness: Overexposed (>1,300), Dark (>1,700)
• Lighting behavior: halos (>1,700), lens flares (>1,700), glass transmission (>3,600), specular reflections (>3,600)