Asset Preprocessing Overview¶

Asset preprocessing (navarena-forge / navarena_forge) is a modular pipeline that converts raw 3D Gaussian Splatting scenes into standardized embodied navigation assets. It transforms PLY point clouds into coordinate-aligned scenes, occupancy grids, navigable region masks, and optional compressed formats for use by the data generator and evaluation framework.

Project Scope¶

• Input: Raw 3DGS PLY point clouds (from InteriorGS, ScanNet++, SceneSplat, etc.)
• Output: Standardized scene directories in the V1 unified asset format

Core Architecture¶

flowchart TB
    subgraph Pipeline[Pipeline]
        S1[coordinate_normalize]
        S2[pcd_to_map]
        S3[valid_region_estimate]
        S4[compress_ply]

        S1 --> S2 --> S3 --> S4
    end

    subgraph Steps[Step Registry]
        Registry[StepRegistry]
        S1 -.register.-> Registry
        S2 -.register.-> Registry
        S3 -.register.-> Registry
        S4 -.register.-> Registry
    end

    subgraph Output[Output Assets]
        Manifest[manifest.json]
        Aligned[aligned.ply]
        NavMap[nav_map.pgm]
        NavMask[nav_mask.png]
        Splat[compressed.splat]
    end

    Pipeline --> Output

Design Patterns¶

1. Template Method - ProcessingStep base class enforces validate → process → save flow
2. Pipeline - Chains steps, manages intermediate artifacts and error handling
3. Registry - Steps auto-register via @StepRegistry.register("name") decorator
4. Strategy - Algorithms (RANSAC, Alpha Shapes) are decoupled from steps

Adding New Steps¶

To add a step:

1. Create a class inheriting ProcessingStep under steps/
2. Register with @StepRegistry.register("my_step")
3. Import in steps/__init__.py and add to pipeline.yaml

V1 Unified Asset Format¶

Each processed scene directory follows this structure:

{dataset}/{scene_id}/
├── manifest.json          # Metadata and provenance (schema v1.0)
├── source.ply             # Original 3DGS point cloud
├── aligned.ply           # Coordinate-normalized PLY
├── nav_map.pgm           # 2D occupancy grid (ROS-compatible)
├── nav_map.yaml           # Map config
├── nav_mask.png           # Navigable region mask
├── compressed.splat       # Compact binary (32 bytes/Gaussian, optional)
├── labels.json           # Semantic labels (optional)
└── structure.json        # Floor plan (optional)

Scene ID Generation¶

Scene ID is the first 8 hex characters of SHA-256("{dataset}:{original_name}"):

• The same (original_name, dataset) always yields the same ID
• Example: InteriorGS:scene_001 → a1b2c3d4

Fixed Filenames¶

The above filenames are fixed; they are not prefixed by scene_id, for consistent downstream parsing.

See also: Pipeline Steps · Configuration · CLI Commands · Web Viewer