duinodu / duinodu.github.io Goto Github PK

Problem

Main stream methods, such as VoxelNet, PointPillar and SECOND, are based on pre-defined anchor. Anchor requires prior knowledge about object. Besides, IOU matching strategy presents an imbalance problem.

Idea

anchor-free

Pipeline

Hotspot is center used in CenterNet. Besides anchor-free, they add quandrant classification.
The idea behind quandrant classification is that they hope network to learn the inherent and invariant object part relation.

Inspiration

Anchor-free is the future. The more automatic, the better result.

Problem

BEV keeps object seperated and preserve physical dimensions, but pointcloud in bev is sparse and have highly variable point density.

Idea

Two main idea:

• Multi view (bev and front-view) fusion.
• Dynamic voxelization.

Pipeline

TBD

Problem

A large number of objects are under heavy occlusions. Useful information for occluded objects is limited. This paper focus on spatial relationship between close-by objects.

Idea

The key idea is to optimize the predicted 3D locations of objects guided by uncertainty-aware spatial contraints.

• pair prediction
• aleatoric uncertainty

Pipeline

The third branch is to predict world distance between two close objects. And for data uncertainty, they use aleatoric uncertainty to make loss more robust to noisy input in regression task. See below for more details.

Inspire

Spatial contraints add more supervision signals to network. For a 3D world with much geometry information, how to add useful contraint to network is an open problem.