|
Algorithm 1 SGMR-LPR: Forward Pass and Joint Training Procedure |
Input: Raw LiDAR point cloud P; Ground-truth semantic BEV (if training phase)
Parameters: Movable class set ; Attention balance coefficient ; Loss weights , ,
Output:
-normalized global place descriptor
Stage 1: Unified Representation Learning
-
1:
Generate BEV image from raw point cloud P ▹Equation (1)
-
2:
▹: high-level BEV features, : semantic logits
Stage 2: Novel Semantic-Guided Feature Modulation (Ours)
— Probabilistic Movable Object Masking (PMOM) —
-
3:
-
4:
▹ Soft uncertainty-aware masking, Equation (4)
— Movable-suppressed Channel-Spatial Attention (MSCS) —
-
5:
-
6:
-
7:
▹ Explicit dynamic suppression, Equation (5)
Stage 3: Global Descriptor Aggregation
-
8:
▹ Spatial & channel dimension adjustment
-
9:
Stage 4: Multi-Task Joint Optimization (Training phase only)
-
10:
if training then
-
11:
▹ Given descriptors extracted from a training tuple
-
12:
-
13:
▹ Equations (10)–(12)
-
14:
▹ Equation (8)
-
15:
Update model parameters by minimizing
-
16:
end if
-
17:
return
|