Marine Debris Roadmap v5

FUTURE · RESOURCE / ACCURACY PARETO GATE

Pareto Analysis + Independent Candidate Pools

Use the full shared audit to form separate Tuning, Ensemble, and Internal-MoE pools. Do not reuse one raw-mAP top-three shortlist for every downstream experiment.

MULTI-OBJECTIVENO MASTER SCOREINDEPENDENT POOLS

Post-baseline decision structure

ALL ACCEPTED CANONICAL MODELS
FULL SHARED AUDIT
RESOURCE / ACCURACY PARETO ANALYSIS
Independent candidate pools
TUNING · ≈4–6

High quality, weak-slice strength, Pareto efficiency, untapped convergence, or architecture-specific opportunity.

ENSEMBLE · ≈2–4

Strong individual quality plus empirically complementary errors.

INTERNAL-MoE · ≈1–2

Good baseline plus measured headroom, expertizable structure, and stable training behavior.

TUNING POOL

Evidence-qualified optimization

Expected ≈4–6 initial models. Depth may differ by model; a smaller subset can receive deeper optimization after early targeted experiments.

ENSEMBLE POOL

Quality + complementarity

Expected ≈2–4 models. Large models are allowed; predictions can be generated independently without simultaneous GPU residency.

INTERNAL-MoE POOL

Resource-aware substrate

Expected ≈1–2 models. The largest or highest-mAP detector is not automatically the best expertization target.

Tiny-object bottleneck

Test resolution, native multiscale, crop/tiling, or scale-aware augmentation.

Long-tail bottleneck

Test repeat-factor or class-aware sampling, class-aware augmentation, or loss weighting.

Crowding bottleneck

Test query count, resolution, matching settings, or crop strategy.

Localization bottleneck

Test regression-specific knobs, resolution, or geometry augmentation.