| Aspect | V2X-Seq-TFD + ego | V2X-Seq-TFD-Infra + ego |
|---|---|---|
| Data source | Vehicle CSV + Infra CSV | Infra CSV only |
| Scene origin | AV (ego vehicle at t=49) | TARGET_AGENT (infra target at t=49) |
| Total actors | N₁ (vehicle) + N₂ (infra) | N (all infra-view) |
| ego_mask | True for vehicle-view only | All True |
| road_mask | True for infra-view only | All False |
| Actors predicted | Only vehicle-view (N₁) | All actors (N) |
| V2X edge data | Real matching labels | Empty |
| Edge loss | Cross-view matching | Zero (skipped) |
The key insight: With V2X-Seq-TFD, the ego_mask genuinely distinguishes vehicle vs. infrastructure actors. With V2X-Seq-TFD-Infra, the ego_mask is a dummy mask (all True) because there's only one view. So cooperation='ego' means "use ego_mask", but what that mask contains depends entirely on which dataset you use.