Model Card for PATT-Net

Model Details

• Architecture: ProtectedTransferVAD (dual-branch NativeTemporalBackbone)
• Base Feature Extractor: OpenAI CLIP ViT-L/14 (768-dim)
• Temporal Modeling: Bi-directional LSTM with Masked Temporal Dilated Convolutions
• Intended Use: Video Anomaly Detection (VAD) research and controlled ablation studies.

Training Data

• Pre-training (Temporal Branch): PreVAD dataset (35,279 videos, distinct from XD-Violence).
• Target Fine-tuning: UCF-Crime dataset.

Evaluation and Validation

• Metric: Frame-level AUC and AP, evaluated via center-crop (stride-16) visual-only features.
• Selection Criteria: Best validation video AUC.
• Scientific Limits: The P-B contrast represents the whole system effect (branch mechanism + pretraining). Within this matched frozen-branch design, P-R isolates PreVAD rather than random source initialization. The P-R differences are small and not statistically significant across five seeds (AUC exact two-sided sign-flip p = 0.1875).

Ethical Considerations and Licenses

• The repository code is provided under the MIT License. Third-party dataset and model terms continue to apply to data-derived artifacts.
• Users must comply with the licenses of the underlying third-party datasets (UCF-Crime and PreVAD) and the OpenAI CLIP feature extractor. The repository authors are not responsible for the misuse of these models in surveillance or real-world classification scenarios without appropriate safeguards.