Datasets and Protocols

Datasets and Protocols

RWF-2000

CARE-SIU currently uses RWF-2000 as the primary real-video benchmark.

Prepared data:

• 1,994 readable videos
• binary interaction setting
• fixed train/validation/test manifest
• five-seed repeated experiments

The temporal benchmark is the current primary source of real-data evidence.

---

CARE-Synth-XL — legacy

The original synthetic dataset contains:

• 10,000 clips;
• 24 interaction classes;
• RGB;
• audio;
• pose;
• trajectory;
• corruption/severity metadata;
• client and round metadata.

The dataset is retained for audit purposes, but its pose/trajectory modality experiments are invalidated because class-specific motion templates crossed train/validation/test.

---

CARE-Synth-XL-v2 — required protocol

The redesigned generator should include:

template_id
variant_id
subject_profile_id
environment_id
camera_profile_id
generation_seed
motion_seed
client_id
round_id
split

Split unit

Splitting occurs by template_id, not by generated clip.

template T001 -> train only
template T002 -> validation only
template T003 -> test only

All variants of the same template remain in one split.

Leakage gate

Future synthetic training must fail if:

• a template crosses splits;
• an exact pose hash crosses splits;
• an exact trajectory hash crosses splits;
• per-class template diversity is below threshold;
• required fields are absent;
• modality uniqueness is implausibly low.

---

Test-set discipline

The test set is not used for:

• hyperparameter tuning;
• early stopping;
• threshold optimization;
• selecting a preferred seed;
• deciding which experimental variant to report.