Learning-Deep-Learning

Occlusion-Net: 2D/3D Occluded Keypoint Localization Using Graph Networks

September 2019

tl;dr: Using GNN to reason position of occluded points using multi-view geometry (trifocal tensor),

Overall impression

Occlusion-Net’s builds on the strong baseline of keypoint detection of Mask RCNN and is better in two ways: it can reason the location of the occluded points better, and it classifies whether a point is occluded much more accurately.

The dataset CarFusion should have the same car annotated in different view points and the correspondence is annotated too. Both visible and occluded keypoints are annotated but only visible keypoints are used for training – the occluded ones are used for evaluation only.

Key ideas

Technical details

Notes