Skip to main content

Phase 2: Scene Decomposition (Asset Strategy & Branching)

WORKFLOWADVANCED
📖10 min read

Introduction

Scene decomposition involves isolating, categorizing, and preparing every individual element within the frame for reconstruction. This phase turns a complex scene into a structured inventory of assets ready for object and environmental reconstruction.


Inventory

Identify and categorize all static and dynamic entities.

Intelligence

Gather ground truth data, blueprints, and specs.

Methodology

Determine the best creation strategy for each asset.


Asset Identification (Inventory)

  • Macro-Static (Infrastructure): Buildings
  • Micro-Static (Furniture): Street signs, benches, trees, trash cans.
  • Dynamic Entities (Movers):
    • Biometric: Humans, crowds, animals (Agency).
    • Kinematic: Vehicles, machinery (Physics).
  • Ephemeral Events: Explosions, smoke plumes, fire.
  • Environment: Sun position, weather conditions (wet roads/puddles), debris.

Naming & Organization

Assign a unique ID to every item in the Inventory.

Don't use naming conventions that are subject to change (No Car_Lada_Niva_, but Car_Sedan_White). Pick categories, classifications or a taxonomy and visual aspects that are static (a car doesn't change from black to blue, nor does it change from Sedan to Hatchback).

Naming Convention: Category_Description_ID (e.g., MACRO_Building_TownHall_01 or DYN_Car_RedSedan_02).

Folder Structure: Create a directory for each asset containing two sub-folders:

  • /Reference (Blueprints, OSINT photos).
  • /Data (Measurements, specs).

Intelligence (Reference Gathering)

Try to find "Ground Truth" data for each item in your Inventory.

OSINT / Specs: Locate blueprints, satellite maps, vehicle manuals, or weather reports.

Dimensional Verification: Confirm the height, width, and length of objects. Example: Don't guess the height of the lamp post; find the manufacturer spec or measure it relative to a known object.

Alternative Angles: Find existing footage (Street View, News, Social Media) that shows the object from sides not visible in your primary evidence.


Establish Methodology per object

StrategyDefinitionBest Use Case
SOURCELibrary AssetsMicro-Static & Generic Kinematic. Standard objects (Trash cans, Oak trees, generic Ford Focus) where the specific scratches/dents don't matter. Do not reinvent the wheel.
GENERATEAutomated Creation, ProceduralMacro-Static & Environment. Complex terrain, debris fields, or large city blocks. Use Photogrammetry (Drone/Google Earth) or Scatter tools (Gravel, Grass).
MODELHand-CraftedKey Evidence. The specific building where the event happened, or a unique vehicle with specific damage. These must be built vertex-by-vertex to match the reference perfectly.
SIMULATEPhysics-BasedEphemeral Events. Explosions, fire, smoke. These cannot be modeled; they must be calculated using fluid dynamics based on the visual evidence.

Evidence Assessment

What source material do you have? This determines which reconstruction techniques are possible.

Visual Evidence Density:

  • Sparse (1-5 viewpoints): Limited angles, use camera matching/perspective projection
  • Medium (6-30 viewpoints): Structure-from-Motion (SfM) photogrammetry feasible
  • Dense (30+ viewpoints): Full photogrammetry or Gaussian Splatting possible

Visual Evidence Quality:

  • High: DSLR photos, stabilized video, minimal compression
  • Medium: Smartphone footage, CCTV with moderate resolution
  • Low: Compressed social media video, long-distance footage, heavy motion blur

Temporal Coverage:

  • Single moment: Static reconstruction
  • Discrete events: Keyframe animation between known states
  • Continuous: Full motion tracking required

Supplementary Data Availability:

  • Geospatial: GPS coordinates, satellite imagery, maps, blueprints available?
  • Technical specs: Vehicle dimensions, building plans, object specifications?
  • OSINT: Can you find additional angles from social media, news footage, Street View?

→ Output: Identify which reconstruction techniques are viable (photogrammetry vs. modeling vs. camera matching)


Summary

Scene decomposition breaks the reconstruction problem down into manageable parts. By creating a detailed inventory and assigning a specific creation methodology to each asset, you ensure that effort is focused where it matters most (Key Evidence) while efficiently handling background elements.

Key Takeaways:

  • Inventory: List everything before you build anything.
  • Naming: Use consistent, descriptive naming conventions.
  • Ground Truth: Verify dimensions with external data whenever possible.
  • Strategy: Choose the right tool (Model, Source, Generate, Simulate) for each object.

Documentation for Methodology

What to document during Scene Decomposition

In Your Method Section

"The scene was decomposed into [Number] distinct assets. Key evidence, such as [Object A], was modeled manually based on [Reference]. Contextual elements were sourced from [Library Name] or generated procedurally."

In Your Decision Log

Record the following:

  • The source of dimensions for each key asset (e.g., "Vehicle dimensions from 2018 Manufacturer Manual").
  • Why a specific modeling strategy was chosen (e.g., "Used photogrammetry for the statue due to complex geometry and available drone footage").
  • Any assumptions made about object scale where ground truth was unavailable.

Verification

"Modeled assets were overlaid on source imagery to verify silhouette alignment and scale."


Further Resources: