Segmenting unknown objects from generic scenes is one of the ambitious and elusive goals of computer vision. With the recent introduction of cheap and powerful 3D sensors (such as the Microsoft Kinect or Asus XtionPRO) which deliver a dense point cloud plus color for almost any indoor scene, a renewed interest in 3D methods holds the promise to push the envelope slightly further. The Object Segmentation Database provides RGBD data in several subcategories to enable evaluation of object segmentation approaches. The database contains currently 111 entries, all providing the RGBD image, the color image and ground truth annotation.
Richtsfeld Andreas – ari(at)acin.tuwien.ac.at
Trainingsset:
Boxes (0-16)
Stacked Boxes (17-24)
Occluded Objects (25-32)
Cylindric Objects (33-44)
Testset:
Boxes (0-15)
Stacked Boxes (16-23)
Occluded Objects (24-30)
Cylindric Objects (31-42)
Mixed Objects (43-54)
Complex Scene (55-65)
OSD.tar.gz – OSD.zip
OSD-0.2.tar.gz – OSD-0.2.zip (Update: Corrected annotation for test25 and test26)
Depth images: OSD-0.2-depth.tar.gz – OSD-0.2-depth.zip
Changes in Version 0.2:
– Higher accuracy of ground truth (based on color image, instead of depth image)
– Merged point cloud and annotation in pcl-style as cloud of PointXYZRGBL
A database with simpler, freestanding objects is provided by Willow Garage. Have a look to Willow Garage Dataset to find out more.
Boxes
Stacked Boxes
Occluded Objects
Cylindric Objects
Mixed Objects
Complex Scenes