PointPROP

Parallel Ray Occlusion Prediction for Point Clouds

Overview

PointPROP is a deep learning–based surrogate model for point-in-time solar ray occlusion directly on urban LiDAR point clouds.
Given:

• a context point cloud pts_context ∈ ℝ^{N×3},
• a set of test points pts_test ∈ ℝ^{M×3}, and
• a solar ray direction ray ∈ ℝ^3,

the model predicts whether each test point is occluded or visible along that ray.

Formally, we learn a binary classifier:

$$ f(\text{pts}_{\text{context}},, \text{pts}_{\text{test}},, \text{ray}) \in {0,1}^M $$

Motivation

PointPROP aims to learn a neural surrogate that directly predicts ray occlusion from raw point clouds as motivated by point-in-time solar occulsion testing at urban scale.

Key Ideas

• Treat ray occlusion as a physics-governed, geometry-dependent function, not a semantic task.
• Explicitly design inductive biases that reflect:
  • local geometric structure,
  • global scene configuration,
  • directional constraints (solar vector).
• Compare multiple architectural families:
  • Global MLP head (PointNet-style pooling)
  • Transformer cross-attention head
  • MLP + local k-NN neighborhood pooling (PointNet++-style)

Data

• Source: Urban-scale LiDAR point clouds, partitioned into fixed-size sub-patches.

• For each sub-patch, we extract:

  1. Context points (surrounding geometry)
  2. Test points (query locations)
  3. Solar ray direction (sampled from a hemispherical sky dome)
  4. Binary occlusion labels (visible = 1, occluded = 0)

• Ground truth generated via a voxelization-based simulation:

  • Point cloud → 3D occupancy grid
  • Ray–voxel intersection to determine occlusion

• Dataset size:

  • 123 sub-patches × 16 rays per patch = 1,968 samples
  • Split: 1,377 train / 393 val / 198 test

Model Architecture

Encoders

• Point cloud encoder

  • RandLA-Net–style encoder (without decoder)
  • Random downsampling + Local Feature Aggregation (LFA)
  • Produces per-point features + downsampled coordinates for context and test clouds.

• Positional encoder (PtPositionEncoder)

  • MLP with optional Fourier positional encoding of 3D coordinates.

• Solar encoder (Solar_Embed)

  • MLP with low-frequency Fourier embedding of the solar direction.

Heads

We implement three heads over the shared latent space:

1. ShadingSurrogateMLP (Global-Only)

   • Fuse solar direction into context features (concat / gated / add).
   • Global mean & max pooling over context → scene descriptor.
   • Broadcast descriptor to test points and pass through MLP to predict occlusion.

2. ShadingSurrogateTransformer (Cross-Attention)

   • Fuse solar into context features as above.
   • Multi-Head Attention:
     • Queries: test-point features
     • Keys/Values: solar-conditioned context features
   • Attention output → MLP → per-point logits.

3. ShadingSurrogateMLPV2 (Global + Local)

   • Global branch as in MLP head.
   • Local branch:
     • For each test point, find k-NN in context.
     • Encode relative positions and neighbor features via shared MLP.
     • Pool (mean + max) over neighbors → local descriptor.
   • Concatenate:
     • test-point feature + local feature + global feature → MLP → logits.

All heads are trained with BCE-with-logits loss.