Indoor scene dataset collection

• Berkeley B3DO
• 849 images.
• 50 categories.
• Bounding box level annotation.
• SUN 3D
• polygonal 2D semantic annotation of total of 10335 RGBD images. (additional 3D polygon for 3D detection is also available, 3D polygonal for room layout annotation.)
• 4943 new images captured.
• 3784 images using (Kinect v2).
• 1159 using (Intel Realsense ).
• 1449 from NYUD v2 (Kinect v1).
• 554 from B3DO (Kinect v1).
• 3389 selected frames from SUN3D videos (Asus Xtion).
• 19 scene-types (includes most of the NYUD v2 scenes).
•  Reference: semantic segmentation different CNN models
• Robot in a room 
• Not available.
• NYUD v1
• 7 scenes-types with full labels.
• 12 object classes to classify.
• total 2347 images.
• NYUD v2 
• 9 dominant scenes-types.
• 37 dominant categories of object classes.
• total 1449 images.
• Cornell RGBD 
• Office and home scene-types
• 28 office scenes (each scene has point cloud annotated bag file)
• 24 home scenes
• separate image with label for each frame is available.
• COLD (Ljubljana+Freiburg+Strucken) 
• 5 common scenes (bathroom, kitchen, printer, office, hallway) in all three geo-locations. Additionally other scenes as sequence images.
• RGB only
• karpathy Discovery 
• 58 scenes (like Cornell RGBD) of office+kitchen+printer
• 36 office desks + 7 bookshelves + 4 printers + 3 kitchen coutners + 8 miscellaneous living space scenes.
• Types of objects are different than that of NYUD v2; more object centric.
• 3D meshes representations of the scene from kinect fusion + post-processing.
• paspart_challenge: part segmentation 
•