Skip to main content
. 2026 May 12;26(10):3050. doi: 10.3390/s26103050
Algorithm 1 SGMR-LPR: Forward Pass and Joint Training Procedure
  •    Input: Raw LiDAR point cloud P; Ground-truth semantic BEV BgtS (if training phase)

  •    Parameters: Movable class set Cmov; Attention balance coefficient α=0.5; Loss weights λlt=λseg=1.0, λce=0.4, λdice=0.6

  •    Output: ℓ2-normalized global place descriptor F

  •    Stage 1: Unified Representation Learning

  •   1:

    Generate BEV image BI from raw point cloud P                                          ▹Equation (1)

  •   2:

    F,Z←Backbone(BI)                                        ▹F: high-level BEV features, Z: semantic logits

    Stage 2: Novel Semantic-Guided Feature Modulation (Ours)

    — Probabilistic Movable Object Masking (PMOM) —

  •   3:

    P←Softmax(Z,dim=channel)

  •   4:

    M←Clamp(∑c∈CmovPc,0,1)                                     ▹ Soft uncertainty-aware masking, Equation (4)

    — Movable-suppressed Channel-Spatial Attention (MSCS) —

  •   5:

    Fch←ChannelAttention(F)

  •   6:

    Fsp←SpatialAttention(F)

  •   7:

    F′←αFch+(1−α)(Fsp⊙(1−M))                                       ▹ Explicit dynamic suppression, Equation (5)

    Stage 3: Global Descriptor Aggregation

  •   8:

    Fconv←Conv_Block(F′)                                          ▹ Spatial & channel dimension adjustment

  •   9:

    F←L2_Normalize(NetVLAD(Fconv))

    Stage 4: Multi-Task Joint Optimization (Training phase only)

  • 10:

    if training then

  • 11:

                                                   ▹ Given descriptors Fa,Fp,{Fni} extracted from a training tuple

  • 12:

          Llt←LazyTripletLoss(Fa,Fp,{Fni})

  • 13:

          Lseg←λceLce(P,BgtS)+λdiceLdice(P,BgtS)                                            ▹ Equations (10)–(12)

  • 14:

          L←λltLlt+λsegLseg                                                        ▹ Equation (8)

  • 15:

          Update model parameters by minimizing L

  • 16:

    end if

  • 17:

    return  F