mmvtg

MMVTG: Checkpoints for CCTV-Based Multi-Modal Video Temporal Grounding

This repository hosts the official model checkpoints accompanying our paper:

Cross-Attention-Based Intelligent Video Temporal Grounding for CCTV-Based Crime Identification in Smart Cities

The released checkpoints are intended to support research reproducibility, benchmarking, and future work in multi-modal video understanding and intelligent surveillance.

---

Overview

This work extends the excellent Moment-DETR framework with a reproducible workflow tailored for Multi-Modal Video Temporal Grounding (MMVTG) in CCTV crime investigation scenarios.

Our contributions include:

• Weakly-supervised pretraining using ASR captions
• Fine-tuning on the QVHighlights benchmark
• Inference-time contextual grounding using retrieved FIR (First Information Report) documents
• End-to-end reproducible training and inference workflow
• A full-stack deployment application for interactive inference

This repository contains only the released model checkpoints.

The accompanying GitHub repository contains:

• Source code
• Training scripts
• Colab notebooks
• Deployment application
• Documentation
• Reproducibility guide
• Paper companion resources

---

Available Checkpoints

---

Training Workflow

The released checkpoints correspond to the following pipeline:

ASR Caption Pretraining
        │
        ▼
Fine-tuning on QVHighlights
        │
        ▼
Validation
        │
        ▼
Best Checkpoint Selection
        │
        ▼
Inference
        │
        ▼
Grounded Inference using Retrieved FIR Reports

---

Intended Use

These checkpoints are released for:

• Academic research
• Reproducibility of the accompanying paper
• Benchmarking
• Educational purposes
• Experimental extensions of Moment-DETR

---

Limitations

This work extends the Moment-DETR framework.

The retrieval component is performed only during inference by augmenting the input query with retrieved FIR reports to provide additional contextual information.

This repository does not implement a complete Retrieval-Augmented Generation (RAG) architecture or an end-to-end retrieval-training pipeline.

---

Repository Structure

This Hugging Face repository contains only the released model checkpoints.

The complete project is organized as:

• Research implementation
• Deployment application
• Paper companion repository
• Documentation
• Reproducibility resources

---

Citation

If you use these checkpoints in your research, please cite our paper.

BibTeX will be added once the final publication metadata becomes available.

% Citation coming soon

---

Acknowledgements

This work builds upon the excellent open-source implementation of Moment-DETR developed by Lei et al.

We gratefully acknowledge the original authors for making their implementation publicly available and for providing the foundation upon which this work was developed.

---

Related Resources

---

License

These checkpoints are released under the Apache License 2.0. See the accompanying GitHub repository for full licensing details and usage terms.