BigBIRD: A large-scale 3D database of object instances

摘要: The state of the art in computer vision has rapidly advanced over past decade largely aided by shared image datasets. However, most these datasets tend to consist assorted collections images from web that do not include 3D information or pose information. Furthermore, they target problem object category recognition—whereas solving instance recognition might be sufficient for many robotic tasks. To address issues, we present a highquality, large-scale dataset instances, with accurate calibration every image. We anticipate “solving” this will effectively remove perceptionrelated problems mobile, sensing-based robots. contributions work of: (1) BigBIRD, 100 objects (and growing), composed of, each object, 600 point clouds and high-resolution (12 MP) spanning all views, (2) method jointly calibrating multi-camera system, (3) details our data collection which collects required single under 6 minutes minimal human effort, (4) multiple software components (made available open source), used automate multi-sensor process. All code are at http://rll.eecs.berkeley.edu/bigbird.