COD Map Evaluator

A machine learning system that analyzes Call of Duty minimap images to predict map quality ratings for Hardpoint (HP) and Search & Destroy (SnD) game modes.

This project explores whether a map will play well based purely on its layout geometry, using minimap images as input. It's a research/tooling project focused on map flow, rotations, choke points, and accessibility - not gun balance, player skill, or full 3D simulation.

Current Pipeline: Minimap Image → Walkable Space Extraction → Geometric Metrics → Ridge Regression → Quality Scores (1-10)

---

Project Vision

Long-term Goals:

1. Evaluate existing maps from minimap images
2. Predict whether a map will play well for HP and SnD
3. Break "good map design" into measurable geometry metrics
4. (Future) Generate new map layouts that score highly under those metrics

Philosophy: Start with strong geometric signals first, then introduce ML models. Prioritize interpretability over raw accuracy - if a metric feels unintuitive to a competitive player, it needs revision.

---

Scope (Strict)

Included

• Call of Duty Hardpoint only
• 2D minimap-style images
• Geometry-based evaluation
• Hill-to-hill rotations

Explicitly Excluded (for now)

• Elevation / verticality
• Weapon balance
• Spawn logic (later extension)
• Player behavior modeling
• Full 3D environments

---

Core Idea

Minimap → Walkable Space → Navigation Graph → HP Metrics → Score

Instead of asking “Is this map good?”, we ask:

• How many viable routes reach each hill?
• How severe are choke points?
• How healthy are rotations between hills?
• Is control too centralized?

These questions are turned into quantitative metrics.

---

Outputs (Label Schema)

Each map is evaluated using the following scores (0–1):

• rotation_quality – health of hill-to-hill rotations
• hill_accessibility – number & width of viable hill entrances
• choke_severity – dominance of narrow choke points
• overall_hp_score – weighted aggregate HP quality

Labels are initially human-assigned and later learned.

---

Data

Input

• Grayscale minimap images (normalized resolution)
• Manual hill center annotations (initial MVP)

data/
  maps/
    hotel_hp.png
    skidrow_hp.png
    invasion_hp.png
  labels.csv

Assumptions

• Light regions ≈ walkable space
• Dark regions ≈ walls / blocked space
• Hills are fixed points of interest

---

Pipeline

1. Image Normalization

• Resize to fixed resolution (e.g. 256×256)
• Convert to grayscale

2. Walkable Space Extraction

• Threshold-based segmentation
• Morphological cleanup
• Binary walkable mask

3. Navigation Graph Construction

• Downsample walkable mask
• Nodes = walkable cells
• Edges = adjacency

4. Metric Computation (Per Hill)

• Shortest paths
• Path diversity
• Betweenness centrality
• Entrance width estimation

5. Hardpoint Scoring

• Hand-designed weighted score
• Later replaced / augmented by ML

---

Evaluation Philosophy

This project prioritizes:

• Interpretability over raw accuracy
• Geometry-first reasoning before ML
• Simple heuristics that mirror how players talk about maps

If a metric feels unintuitive to a competitive player, it probably needs revision.

---

Machine Learning (Later Phase)

Once geometric metrics are reliable:

• CNN predicts HP scores directly from images
• Hybrid models combine CNN + graph features
• Models are used as evaluators for generation

ML is a tool, not the foundation.

---

Map Generation (Future Work)

Generation is constraint-driven:

• Lane count
• Symmetry type
• Choke density targets
• Hill spacing

Layouts are:

1. Generated procedurally
2. Scored by evaluator
3. Mutated / evolved

This forms an evolutionary map design loop.

---

Project Status

• Collect minimaps
• Walkable space extraction
• Hill annotation
• Graph construction
• Metric validation
• Baseline CNN
• Procedural generation

---

Why This Project Exists

Competitive FPS maps live or die on geometry.

This project attempts to:

• Formalize intuitive map design knowledge
• Provide tooling for designers & analysts
• Explore AI-assisted multiplayer level design

---