Abstract: Notes for Architecture and building blocks for the RandLA which is implemented in the experiment

Building Blocks

1. Local feature aggregation 1.Local Spatial encoding (LocSE) –> 3+d: xyz position + other per-point features (e.g. RGB or other intermediate learned features)

      Objective
      ​	**Encode xyz coordinates of all neighbouring points for each point to aware of relative spatial locations.**
      Steps:
       1. Finding neighbouring points
             for $ith$ point, gather neighbouring points by $KNN$ for efficiency, based on Euclidean distances.
      2. Relative Point Position encoding
           concatenate the xyz positions of ith points, neighbour k of ith points, element-wise subtraction, Euclidean distance of two points
      3. Point Feature Augmentation
           Concatenate relative point position with corresponding point features f.
   

2.Attentive pooling Objective ​ Aggregate the set of neighbouring point features f, use attention mechanism to automatically learn important local features. Steps: 1.Computing Attention Scores 2.Weighted Summation 3.dilated residual block

 ​