manhdungcr7/cuenet-rwf2000

CUE-Net: Violence Detection Model

Model Description

CUE-Net (CLIP-based UniFormerV2 Enhanced Network) là mô hình phát hiện bạo lực từ video giám sát, được huấn luyện trên bộ dữ liệu RWF-2000.

Architecture

• Backbone: CLIP ViT-L/14-336
• Framework: UniFormerV2
• Input: 64 frames × 336 × 336
• Classes: 2 (Fight, NonFight)
• Parameters: ~354M

Performance

Metric	Score
Accuracy	89.50%
F1-Score	89.48%

Usage

from slowfast.models.build import build_model
from slowfast.config.defaults import get_cfg

cfg = get_cfg()
cfg.merge_from_file("config.yaml")
model = build_model(cfg)

# Load checkpoint
checkpoint = torch.load("cuenet_rwf2000_epoch51.pyth")
model.load_state_dict(checkpoint["model_state"])

Training Details

• Optimizer: AdamW
• Learning rate: 4e-4
• Epochs: 51
• Batch size: 2-4