Struct RerunObserver 

pub struct RerunObserver { /* private fields */ }

Rerun observer for real-time optimization visualization.

This observer logs comprehensive optimization data to Rerun for interactive visualization and debugging. It implements the OptObserver trait, enabling clean integration with any optimizer through the observer pattern.

What Gets Visualized

• Time series: Cost, gradient norm, damping (LM), step norm, step quality
• Matrices: Sparse Hessian (downsampled heat map), gradient vector
• Poses: SE2/SE3 manifold states updated each iteration
• 3D Landmarks: Rn variables with dimension=3 visualized as point clouds
• Status: Convergence information

Observer Pattern Benefits

• Decoupled from optimizer internals
• Can be combined with other observers (CSV, metrics, etc.)
• No #[cfg(feature = "visualization")] scattered through optimizer code
• Easy to enable/disable without changing optimizer logic

Performance

The observer is designed to have minimal overhead:

• Matrix visualizations use downsampling (100×100 for Hessian)
• Rerun logging is asynchronous
• When disabled, is_enabled() returns false immediately
• 3D landmarks are batch-logged as a single point cloud for efficiency

Pose Convention Support

For bundle adjustment (BAL datasets), camera poses are stored as world-to-camera transforms (T_wc). Set invert_camera_poses = true to display cameras correctly by converting to camera-to-world (T_cw) convention for Rerun visualization.

Implementations

impl RerunObserver

pub fn new(enabled: bool) -> ObserverResult<Self>

Create a new Rerun observer.

Arguments

• enabled - Whether to enable visualization

Returns

A new observer instance that spawns a Rerun viewer (or saves to file if viewer unavailable).

Examples

use apex_solver::observers::RerunObserver;

let observer = RerunObserver::new(true)?;

pub fn new_with_options( enabled: bool, save_path: Option<&str>, ) -> ObserverResult<Self>

Create a new Rerun observer with file save option.

Arguments

• enabled - Whether to enable visualization
• save_path - Optional path to save recording to file instead of spawning viewer

Examples

use apex_solver::observers::RerunObserver;

// Save to file
let observer = RerunObserver::new_with_options(true, Some("opt.rrd"))?;

// Spawn live viewer
let observer2 = RerunObserver::new_with_options(true, None)?;

pub fn with_config( enabled: bool, save_path: Option<&str>, config: VisualizationConfig, ) -> ObserverResult<Self>

Create a new Rerun observer with full configuration.

This is the primary constructor for full control over visualization.

Arguments

• enabled - Whether to enable visualization
• save_path - Optional path to save recording to file instead of spawning viewer
• config - Visualization configuration

Examples

use apex_solver::observers::{RerunObserver, VisualizationConfig};

let config = VisualizationConfig::new()
    .with_show_cameras(true)
    .with_show_landmarks(false)
    .with_camera_fov(0.8);

let observer = RerunObserver::with_config(true, None, config)?;

pub fn new_for_bundle_adjustment( enabled: bool, save_path: Option<&str>, invert_camera_poses: bool, ) -> ObserverResult<Self>

Create a new Rerun observer configured for bundle adjustment.

This constructor is designed for bundle adjustment / structure-from-motion problems where camera poses are stored in world-to-camera convention (T_wc) but need to be displayed in camera-to-world convention (T_cw).

Arguments

• enabled - Whether to enable visualization
• save_path - Optional path to save recording to file instead of spawning viewer
• invert_camera_poses - If true, invert SE3 poses before logging (T_wc -> T_cw)

Use Cases

• Pose graph optimization: Use invert_camera_poses = false (poses are already T_cw)
• Bundle adjustment (BAL): Use invert_camera_poses = true (BAL stores T_wc)

Examples

use apex_solver::observers::RerunObserver;

// For bundle adjustment with BAL datasets (world-to-camera poses)
let observer = RerunObserver::new_for_bundle_adjustment(true, None, true)?;

// For pose graph optimization (camera-to-world poses)
let observer = RerunObserver::new_for_bundle_adjustment(true, None, false)?;

pub fn config(&self) -> &VisualizationConfig

Get the current visualization configuration.

pub fn is_enabled(&self) -> bool

Check if visualization is enabled and active.

pub fn set_iteration_metrics( &self, cost: f64, gradient_norm: f64, damping: Option<f64>, step_norm: f64, step_quality: Option<f64>, )

Set iteration metrics for the next on_step call.

This method should be called by optimizers before notifying observers to provide context like cost, gradient norm, damping, etc.

Arguments

• cost - Current cost value
• gradient_norm - L2 norm of gradient
• damping - Current damping parameter (LM-specific, use None for GN/DogLeg)
• step_norm - L2 norm of parameter update
• step_quality - Step quality metric ρ (actual vs predicted reduction)

Examples

observer.set_iteration_metrics(
    1.234,      // cost
    0.056,      // gradient_norm
    Some(0.01), // damping (LM only)
    0.023,      // step_norm
    Some(0.95), // step_quality
);

pub fn set_matrix_data( &self, hessian: Option<SparseColMat<usize, f64>>, gradient: Option<Mat<f64>>, )

Set matrix data (Hessian and gradient) for visualization.

This should be called before on_step if you want to visualize matrices.

Arguments

• hessian - Optional sparse Hessian matrix (J^T J)
• gradient - Optional gradient vector (J^T r)

pub fn log_initial_graph(&self, graph: &Graph, scale: f32) -> ObserverResult<()>

Log the initial graph structure before optimization.

This should be called once before optimization starts to visualize the initial configuration.

Arguments

• graph - The graph structure loaded from G2O file
• scale - Scale factor for visualization

pub fn log_convergence(&self, status: &str) -> ObserverResult<()>

Log convergence status and final summary.

Call this after optimization completes.

Arguments

• status - Convergence status message

pub fn log_initial_ba_state(&self, problem: &Problem) -> ObserverResult<()>

Log initial bundle adjustment state before optimization.

This method visualizes the initial camera poses and 3D landmarks before optimization begins, allowing comparison with optimized results.

Arguments

• problem - The optimization problem containing the initial variables

Examples

use apex_solver::observers::RerunObserver;
use apex_solver::core::problem::Problem;

let observer = RerunObserver::new_for_bundle_adjustment(true, None, true)?;
let mut problem = Problem::new(apex_solver::linalg::JacobianMode::Sparse);
// ... add variables and factors ...

observer.log_initial_ba_state(&problem)?;

Trait Implementations

impl Default for RerunObserver

fn default() -> Self

Returns the “default value” for a type.

impl OptObserver for RerunObserver

fn on_step( &self, values: &SlotMap<VarKey, Box<dyn ManifoldVariable>>, iteration: usize, )

Called at each optimization iteration.

This logs all visualization data to Rerun, including:

• Time series plots (cost, gradient, damping, step quality)
• Matrix visualizations (Hessian, gradient) if set via set_matrix_data
• Manifold states (SE2/SE3 poses)

In InitialAndFinal mode, this method only logs scalar metrics (plots) during intermediate iterations. The full manifold state is logged at iteration 0 (initial) and in on_optimization_complete (final).

Arguments

• values - Current variable values (manifold states)
• iteration - Current iteration number

fn on_optimization_complete( &self, values: &SlotMap<VarKey, Box<dyn ManifoldVariable>>, iterations: usize, )

Called when optimization completes.

In InitialAndFinal mode, this logs the final optimized state. In Iterative mode, the final state was already logged via on_step.

Arguments

• values - Final optimized variable values
• iterations - Total number of iterations performed

fn set_iteration_metrics( &self, _cost: f64, _gradient_norm: f64, _damping: Option<f64>, _step_norm: f64, _step_quality: Option<f64>, )

Set iteration metrics for visualization and monitoring.

fn set_matrix_data( &self, _hessian: Option<SparseColMat<usize, f64>>, _gradient: Option<Mat<f64>>, )

Set matrix data for advanced visualization.