Depth to Z Fusion

Name: “depth_to_z_fusion”

Description

Fuse monocular depth predictions onto stereo Z through spatial smoothing. This application improves Z estimates by leveraging additional depth information (e.g., from trained depth models) while preserving attachment to the original stereo triangulation. The fusion solves an energy minimization problem that balances edge-aware smoothness with data fidelity.

This application runs after point cloud outlier removal steps and before point cloud refinement. It operates on the per-tile surface at final resolution, producing a fitted depth map on the domain defined by depth input coverage.

Configuration

Name	Description	Type	Available value	Default value	Required
method	Method for depth to Z fusion	string	“anisotropic”	“anisotropic”	No
activated	Run this application (if false, skip processing)	boolean, string	“auto”, true, false	“auto”	No
save_intermediate_data	Save fitted depth map and diagnostic layers as TIF	boolean		false	No

Note

• The activated parameter accepts "auto" and is used when the monocular/global preprocessing decides the value automatically.

• When activated is set to false, the application is skipped entirely.

If method is anisotropic:

Name	Description	Type	Available value	Default value	Required
lambda_data	Weight balancing data attachment (Z) vs. smoothness; higher -> closer to original Z	float	>= 0	0.02	No
depth_sigma	Edge-preservation scale. Edges sharper than this are preserved; None -> per tile auto-estimate (recommended)	float	> 0 or None	None	No
depth_sigma_auto	Auto-estimate depth_sigma from local depth statistics if depth_sigma is None	boolean		true	No
detail_scale	Gain on the depth-slope relief (1.0 = match depth slopes at full strength, >1 = stronger, <1 = weaker)	float	>= 0	1.0	No
cross_tile_weight	Weight for cross-tile boundary constraints (higher -> smoother tile boundaries)	float	>= 0	1.0	No
fill_values	None = keep Z unchanged, “invalid” = fill invalid Z only, “all” = replace all Z values	string	None, “invalid”, “all”	“invalid”	No
iterations	Maximum Jacobi solver iterations	integer	> 0	50	No
tol	Convergence tolerance for Jacobi solver	float	> 0	1.0e-4	No

Outputs

The fused surface is saved with the following bands:

• fit_depth_map.tif: Fitted Z from depth-guided fusion (defined on depth domain, with Z anchor subset)

• depth_to_z_residual.tif: Difference (Z - fitted Z) at Z anchor points; diagnostic only

• depth_to_z_weight_map.tif: Per-pixel smoothness weight from edge structure; diagnostic only

• filling.tif: Filling provenance map (EPI_FILLING). Depth-to-Z updates the depth_to_z filling band where Z values were replaced.

By default (fill_values: "invalid"), valid stereo Z values are preserved and only invalid Z values are filled from fit_depth_map. If fill_values: "all", all Z values are replaced by fitted values. If fill_values: null, no replacement is applied.

Example

YAML JSON Copy

---
applications:
  depth_to_z_fusion:
    method: anisotropic
    lambda_data: 0.01
    depth_sigma: null
    depth_sigma_auto: true
    detail_scale: 1.0
    cross_tile_weight: 1.0
    fill_values: invalid
    iterations: 50
    tol: 1.0e-4
    save_intermediate_data: false

{
    "applications": {
        "depth_to_z_fusion": {
            "method": "anisotropic",
            "lambda_data": 0.01,
            "depth_sigma": null,
            "depth_sigma_auto": true,
            "detail_scale": 1.0,
            "cross_tile_weight": 1.0,
            "fill_values": "invalid",
            "iterations": 50,
            "tol": 0.0001,
            "save_intermediate_data": false
        }
    }
}