Real-Time Surface Reconstruction (Incremental BPA)

Author: Hidden for double blind submission

This repository provides an incremental Ball Pivoting Algorithm (BPA) pipeline for real-time and offline surface reconstruction from point clouds.

Overview

The project includes two primary execution paths:

• Offline C++ executable with Open3D visualization in core_ws
• ROS wrapper for online processing from topics or bag files in catkin_ws

This implementation is based on:

Digne, Julie. "An analysis and implementation of a parallel ball pivoting algorithm." Image Processing On Line 4 (2014): 149-168.

Requirements

• CMake 3.11 or newer
• C++20-compatible compiler toolchain
• ROS (required only for catkin_ws)

Note: Dependencies are automatically downloaded and extracted during the build process—no manual setup required.

Third-Party Libraries

Library	Version	Source
yaml-cpp	0.8.0	Fetched via CMake FetchContent
Open3D	0.18.0	Pre-built binary fetched via CMake FetchContent

Quick Start

Offline (recommended for first run)

1. Download sample TXT data from Google Drive sample data into 01-data/01-input.

2. Build:

   cd core_ws
   ./rebuild.sh
   cd ..

3. Run one of the following:

   • Point cloud with normals:

     ./core_ws/build/core_ibpa -c ./config_demo_robot_normal_points.yaml

   • Point cloud without normals (uses sensor pose):

     ./core_ws/build/core_ibpa -c ./config_demo_robot_pose_points.yaml

Online (ROS)

1. Download sample ROS bag data from Google Drive sample data into 01-data/01-input.

2. Start ROS master:

   roscore

3. Play a bag file (example):

   rosbag play -r 1 /path/to/your.bag

4. Build and launch wrapper:

   cd catkin_ws
   ./rebuild.sh

The launch file catkin_ws/src/ros_wrapper_ibpa/launch/ibpa.launch passes config_path to the node. Update it to match your YAML configuration file.

Repository Layout

• core_ws/: C++ library and offline executable (core_ibpa)
• catkin_ws/: ROS wrapper node and launch files
• config*.yaml: Dataset-specific configuration examples
• 01-data/: Input/output data directory

Configuration (YAML)

Minimal annotated example:

# Surface Reconstruction Configuration
radius: 1.0                   # BPA ball radius; also sets octree depth
max_orphan_per_voxel: 1       # Max vertices kept per octree leaf after downsampling
reading_per_batch: 5          # Number of frames fused per batch

main_mesh_policy:
  enabled: true               # Reserved (parsed today, subject to change)
  execute_at_batch: 10        # Activation batch index if enabled is true

random:
  random_device: false        # true = non-deterministic seed
  seed: 42                    # used when random_device is false

seed_triangles_every_batch: false  # Search seed triangulation every batch
show_previous_vertices: false      # Visualize previously seen vertices

# Offline input (core_ws)
input_file: ./01-data/01-input/poses_and_points.txt
pcd_with_normal: false             # true if file contains per-point normals

# Output base path
output_file: ./01-data/02-output/Final_Mesh

# Online input (ROS wrapper)
down_sample_in_ros:
  enabled: true
  max_points: 1000

hole_length: 3.16                  # Boundary length threshold (visualization)

Parameter Notes

• radius: Sets BPA radius and octree depth. Smaller values capture finer detail with higher computational cost.
• max_orphan_per_voxel: Leaf cap used during downsampling. Higher values retain more orphan vertices.
• reading_per_batch: Number of frames fused per iteration (latency vs. stability trade-off).
• seed_triangles_every_batch: Enables seed triangulation search every batch.
• show_previous_vertices: Toggles visualization of previously accumulated vertices.
• down_sample_in_ros: Enables subsampling for incoming ROS point clouds.
• hole_length: Viewer threshold for coloring long boundary edges (red).
• main_mesh_policy: If enabled, keeps only the largest edge-connected mesh after batch N (execute_at_batch). The same effect can be applied interactively by pressing t in the visualization window.

Offline Input File Format

Input files are parsed in blocks delimited by ---.

1. Without normals (pcd_with_normal: false), normals are estimated from pose:

---
pose:
r11 r12 r13 tx
r21 r22 r23 ty
r31 r32 r33 tz
points:
x y z
x y z
...
---

2. With normals (pcd_with_normal: true), each point line contains x y z nx ny nz:

---
points:
x y z nx ny nz
x y z nx ny nz
...
---

Outputs

At program exit, the mesh is written as a PLY file via Open3D:

• Base path: output_file from YAML
• Output pattern: <output_file>_YYYYMMDD_HHMMSS.ply
• Example: 01-data/02-output/Final_Mesh_20250612_153045.ply

ROS Wrapper Details

• Subscribes to:
  • /depth_registered/points (sensor_msgs/PointCloud2)
  • /rovio/odometry (nav_msgs/Odometry)
• TF handling:
  • Attempts to read transform imu_frame -> camera_frame
  • Falls back to a hardcoded extrinsic transform if TF is unavailable. See line 68 in rosIbpaReconstructior.cpp
• Launch file:
  • catkin_ws/src/ros_wrapper_ibpa/launch/ibpa.launch (sets config_path)