One modified detector
A router selects sparse experts inside a single detector. Evaluate accuracy, routing, load balance, VRAM, FLOPs, and stability against the same detector without expert blocks.
A gated ensemble
Independently trained detector families act as experts. This belongs to ensemble evaluation and is not internal MoE.
| Dimension | Required evidence |
|---|---|
| Baseline quality | Project mAP50 plus relevant weak-slice evidence |
| Parameter structure | Total, shared, candidate expertized, and added parameters per expert |
| Optimizer cost | Estimated optimizer-state expansion |
| Memory | Measured baseline training VRAM and remaining headroom |
| Architecture | Clear FFN / MLP expertization points and expertizable parameter fraction |
| Runtime | Activation-memory behavior, throughput, and training stability |
| Scientific value | Plausible specialization benefit; compare vanilla detector vs sparse MoE |