-
3d obj detection(localization + recognition or say bounding box + semantic classification).
-
3d detection apps: vr, ar, autonomous driving, robotics.
-
2d obj detection pop works(faster rcnn, mask rcnn)
-
3d data representations (mesh,volumetric,point cloud, bird's ey view images)
-
current status of 3d detection: heavily rely on 2d-based detectors in various aspects
- voxelize point clouds and apply 3d cnn detector;(high comp cost and fails to leverage sparity in the data)
- images plus 2d image detectors(sacrifice geo details)
- f-pointnet(strictly dependent on the 2d detector, will miss the object entirely if not detected in 2d)
-
what did the author try to accomplish?
-
what were the key elements of this approach?
-
what can u use yourself?
-
imp refs you'd like to cite?