Problem
3D Object Detection
Input: 2d image
Output: 2d detection and segmentation
3d shape of each object
Occlusion and truncation labels
Depth ordering map
Overview
Training using 3d voxel exemplars
A unique 3d voxel pattern representation
Related Work
Basically only do 2d detection bounding boxes. Not much more.
Object Category Recognition
Dataset
KITTI - large dataset of videos of cars driving in challenging image areas
Paper
3D Voxel Exeplars
Description
Occlusion Reasoning
Leverages occlusion reasoning, modeling occlusion by the occluding and occluded objects relationship.
Underlying
intuition is that 1) all the invisible regions of selected
detections shall be explained either by another occluding
object or by image truncation, and 2) visible regions of selected
detections should not overlap with each other. The
model is formulated below.
'mi' is 2D visibility mask composed of three components: 'mv_i' (visible region), 'mo_i' (occluded region), and 'mt_i' (truncated region)
Pipeline Overview
E(D)= Expected Detection Hypothesis (vector of binary values)
New Note
Description