Skip to content

Results

Anubha Parashar edited this page Aug 11, 2026 · 1 revision

Results

1. Reporting policy

CARE-SIU separates:

  • validated real-data results;
  • historical baselines;
  • negative findings;
  • invalidated experiments;
  • pending experiments.

This prevents implementation breadth from being confused with scientific evidence.


2. RWF-2000 temporal benchmark

The main real-data benchmark uses 1,994 readable RWF-2000 videos.

Five random seeds:

1, 7, 21, 42, 84

Main classification comparison

Experiment Macro-F1 ↑ Status
Earlier pooled/weak RGB baseline 0.6661 ± 0.0067 Historical
Temporal R3D-18 real-only 0.8134 ± 0.0121 Validated
Temporal R3D-18 synthetic→real 0.8062 ± 0.0080 Validated transfer experiment

The temporal real-only model improves over the earlier pooled benchmark by approximately:

+0.1473 absolute macro-F1


3. Seed-wise temporal results

Seed Real-only Synthetic→real Δ
1 0.8225 0.8018 −0.0207
7 0.8274 0.7947 −0.0327
21 0.8023 0.8098 +0.0075
42 0.8150 0.8149 −0.00003
84 0.7998 0.8098 +0.0100

Summary:

Statistic Value
Real-only mean 0.8134
Real-only SD 0.0121
Transfer mean 0.8062
Transfer SD 0.0080
Mean Δ −0.0072
Median Δ ≈ −0.00003
Improved seeds 2 / 5

4. Paired statistical analysis

Test Statistic p-value
Paired t-test −0.8599 0.4383
Wilcoxon signed-rank 5.0 0.6250

Interpretation

The current temporal experiment does not support the hypothesis that synthetic pretraining improves real-world classification performance.

This negative result is retained because it constrains the claim and informs the redesign of the synthetic-data pipeline.


5. Calibration and ranking metrics

Approximate aggregate metrics from the completed temporal runs:

Metric Real-only Synthetic→real Better
ECE ↓ ~0.1237 ~0.1164 Transfer
AUROC ↑ ~0.8940 ~0.8901 Real-only
AUPRC ↑ ~0.900 ~0.900 Comparable

The transfer condition therefore shows a small calibration improvement without a macro-F1 improvement.


6. Historical pooled transfer experiment

Before the temporal R3D-18 upgrade, the five-seed pooled/weak experiment produced:

Condition Macro-F1
Real-only 0.6661 ± 0.0067
Synthetic→real 0.6574 ± 0.0252
Mean Δ −0.0087

Statistical tests:

  • paired t-test: p = 0.463470
  • Wilcoxon: p = 0.437500

The temporal encoder substantially improved the real-data representation quality, but synthetic pretraining still did not become a supported classification improvement.


7. Synthetic ablation results — invalidated

The corrected-path synthetic ablation rerun produced near-perfect results for several modalities, including pose-only and trajectory-only.

Those scores are not valid experimental evidence.

Leakage audit

Finding Value
Dataset rows 10,000
Classes 24
Unique pose contents 54
Unique trajectory contents 50
Combined pose/trajectory templates 54
Pose templates in multiple splits 54
Trajectory templates in multiple splits 50
Pose templates used by multiple labels 0
Rows involved in pose cross-split duplication 10,000

Most labels contained only two independent pose templates.

Therefore the model could identify a class-specific template already seen in training.

Status

Synthetic ablation performance:
INVALID FOR GENERALIZATION CLAIMS

The affected results should be retained only as an audit trail.


8. What can currently be claimed?

Supported

  • temporal video modeling substantially improves the real-only RWF-2000 benchmark relative to the earlier pooled representation;
  • five-seed temporal performance is stable around macro-F1 0.81;
  • synthetic pretraining does not significantly improve temporal real-world classification under the completed protocol;
  • calibration shows a small transfer-related improvement;
  • the original synthetic generator contained severe cross-split template leakage, discovered through content hashing and template analysis.

Not yet supported

  • general multimodal superiority over RGB on real-world data;
  • improved real classification from synthetic pretraining;
  • state-of-the-art claims based on the invalidated synthetic ablations;
  • external cross-dataset generalization;
  • deployment-level safety guarantees.

9. Next decisive experiments

  1. Generate leakage-controlled CARE-Synth-XL-v2.
  2. Require template-disjoint train/validation/test splits.
  3. Re-run the 45 modality ablations.
  4. Add a second independent real dataset.
  5. Evaluate missing-modality and corruption robustness.
  6. Quantify selective prediction / review coverage.
  7. Strengthen explanation faithfulness evaluation.