SCNet: Semantic Consistency Networks for 3D Object Detection

The framework uses VoteNet and PointNet++ as the pipeline.

• Semantic Voting: Semantic information is also used as information for each point voting (prediction)
  

1. The two MLP branches complete the voting of normal VoteNet (xyz coordinates and feature features) and Semantic Vote respectively
2. Combines the two branch predictions at each point

• Loss Function
  
  Set hyperparameter weights for eachTask loss is weighted
• Semantic Consistency Mechanism and Loss
  
  as aboveAs shown, take the center of each BBox as the center of the sphere, and set the point within the sphere with a radius of 0.2m to calculate the semantic consistency loss:
  
  pi is the predicted probability of BBox, sj is the semantic information of each query point.
  The model can learn the relationship between geometric information and semantic information, making the prediction of BBOX more accurate.