Back to Maps

Methodology

How the spawn prediction system works

Approach

This system uses behavioral observation by extracting game state from broadcast footage rather than memory reading or code extraction. Everything is derived from what is visible on screen during CDL matches.

The pipeline processes competitive Call of Duty VODs through computer vision to reconstruct game state, then applies machine learning to predict spawn locations from that state.

Data

Training samples109,016 spawn events
Source footage60 competitive VODs
Maps covered9 CDL rotation maps
Events trackedDeaths, spawns, positions, objectives

All data was collected from publicly available tournament broadcasts and creator content under fair use for research purposes.

Key Findings

0.956First to Die, First to Spawn

Death order and spawn order are nearly perfectly correlated. The game processes deaths sequentially, creating predictable spawn queues.

70.5%Spawns Do Not Avoid Crowding

Over 70% of waves show spawn point reuse when measured at the wave level. Multiple teammates can spawn at the exact same location within seconds.

0.59Death Spread Predicts Spawn Spread

When your team dies spread out across the map, you spawn spread out. The system respects team geometry.

91.7%Some Spawns Are Near-Deterministic

Exposure P5 sends 91.7% of spawns to a single point. Certain map-hill combinations are almost perfectly predictable.

Note: reuse rates vary by definition (wave-level vs follower-level), but the pattern is consistent.

Technical Stack

The system combines several ML and CV techniques:

  • Object detection and tracking to follow player positions across frames
  • Event detection to identify deaths and spawns from visual state changes
  • Temporal linking to associate each death with its corresponding spawn
  • Feature engineering across spatial, temporal, and objective-relative signals
  • Tree-based models and ensembles for final prediction

The model predicts to 76 discrete spawn points extracted via clustering of observed spawn locations.

What is Shared

This page presents findings and visual evidence, not the implementation. Extraction heuristics, clustering parameters, and training configuration are intentionally omitted.

The novelty is the insight. Replication requires solving the same problems independently.

Limitations

Pro play bias
CDL teams play spawn-optimized strategies. Patterns may differ in ranked.
Hardpoint only
Other modes (Control, Search) have different spawn logic.
No live integration
This is analysis, not a real-time tool.
Rotation inference
Active hardpoint is approximated, not detected directly.

Future Work

When does it break? The model predicts stable spawns well. Can we predict when spawns will flip, suddenly shifting from one side of the map to another?

Sequence models? Initial experiments with sequence models showed no improvement over point-in-time features. Spawns appear to be largely memoryless, each decision based on current state.

From analysis to action. This is observation, not intervention. Real-time integration would require live capture and sub-second inference.

Work continues.