omni

OMNi - Fine-tuned CLIP4Clip (CCTV Anomaly Detection)

This model is a fine-tuned version of CLIP4Clip, customized for CCTV anomaly detection tasks (smoking, fall, vandalism, arson, theft, etc.).

• Author: TaekyungPark (@Tae-P)
• Base model: ViT-B/32 (OpenAI CLIP)
• Training code: GitHub – Tae-P/clip4clip-omni
• Framework: PyTorch
• Training details: fine-tuned on custom balanced CCTV dataset using 4×Titan RTX GPUs

🧠 Model Description

This model fine-tunes the CLIP4Clip architecture for multimodal video-text retrieval, adapting it to detect and retrieve abnormal behaviors in unmanned CCTV environments.
The model uses a dual-encoder CLIP backbone with an additional cross-attention encoder for fusion.

Key configurations:

• linear_patch: 2D
• sim_header: meanP
• loose_type: True
• cross_num_hidden_layers: 4
• max_words: 20
• max_frames: 16

---

⚙️ How to Use

import torch
from transformers import CLIPModel

model = CLIPModel.from_pretrained("taett/omni")
model.eval()