Optimization Backends

Optimization Backends

English | 中文

Python First

The standalone project uses the Python optimizer as the default backend.

Why:

• simpler installation
• no mandatory ROS / catkin requirement
• consistent integration with the PyQt + Open3D GUI
• validated parity against the legacy C++ backend

Optimize Modes

The Python backend now exposes two solve modes:

• Fast ISAM2
  • default in the GUI
  • intended for repeated working-graph updates during manual editing
• Accurate LM
  • batch-style reference solve
  • useful for parity checks or final confirmation before export

In the current GUI:

• Fast ISAM2 lives in Advanced -> Optimize
• MapVoxel also lives in Advanced and defaults to 0.1 m
• TgtVoxel lives in Registration and defaults to 0.1 m

Legacy C++ Fallback

The C++ backend is still supported as an optional fallback.

Use it when:

• you already have the legacy backend environment
• you want direct parity checks against historical runs
• you need a fallback when Python GTSAM is temporarily unavailable

Parameter Consistency

The Python backend follows the same runtime-parameter precedence used by the legacy optimizer:

1. explicit CLI / GUI options
2. runtime_params.yaml
3. validated offline defaults

The legacy C++ fallback still solves with LM. If it receives --optimize-mode isam2, it reports the mode mismatch and falls back to LM explicitly.

Output Files

Both backends export:

• pose_graph.g2o
• optimized_poses_tum.txt
• pose_graph.png
• manual_loop_report.json

The full output layout is:

• manual_loop_projects/<project_id>/
  • edit-state and resume files
• manual_loop_runs/<run_id>/
  • actual optimizer outputs
• manual_loop_exports/<export_id>/
  • final-export manifest pointing to one selected run

global_map_manual_imu.pcd and trajectory.pcd are built during Export by default so iterative graph editing stays responsive.