MissionGNN – Knowledge‑Graph‑Driven Temporal Reasoning for Video Anomaly Detection

  

MissionGNN: Hierarchical Knowledge‑Graph Reasoning Meets Short‑Term Temporal Context for Surveillance Video Anomaly Detection Sanggeon Yun, Ryozo Masukawa, Minhyoung Na, Mohsen Imani In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV 2025)

Figure 1. The framework for mission-specific knowledge graph generation.

Figure 2. The overall framework for our proposed model utilizing the novel concept of hierarchical graph neural network.

MissionGNN attaches a sensor node (video frame embedding) to a class‑specific knowledge graph (KG) and propagates information through custom GCN layers. Encodings from 13 KGs (one per UCF‑Crime event) are concatenated and passed through a lightweight transformer to yield an (Normal + n events) classifier. A decaying threshold progressively suppresses low‑confidence positives during training, mimicking curriculum learning.

---

📁 Repository Structure

MissionGNN/
├── config.py              # Hyper‑parameters, paths
├── datasets.py            # 30‑frame sliding‑window dataset
├── graph/
│   ├── kg_loader.py       # Load & embed per‑class KGs
│   └── layers.py          # Custom GCNConvTarget
├── models/
│   ├── gcn.py             # Batched KnowledgeGCN
│   ├── temporal.py        # Transformer + prediction head
│   └── missiongnn.py      # End‑to‑end model
├── train.py               # Training loop w/ decaying τ
├── evaluate.py            # Checkpoint evaluation
├── requirements.txt       # Tested library versions
└── README.md              # You are here 🚀

---

🚀 Quickstart

1. Clone and install

git clone https://github.com/c0510gy/MissionGNN.git
cd MissionGNN
pip install -r requirements.txt

2. Prepare data & sub‑graphs

Place UCF‑Crime frame embeddings (*.pt) under ./embeddings/ and per‑class KG files under ./subgraphs/:

subgraphs/
 ├── subgraph_Abuse.txt       # edge list: v->u per line
 ├── keywords_Abuse.txt       # key concept words (one per line)
 ├── subgraph_Arrest.txt
 └── ...

3. Train

python train.py

4. Evaluate best checkpoint

python evaluate.py

Outputs:

• Training / validation loss, mAUC, mAP
• Best model in checkpoints/best.pt

---

🔧 Key Configs (config.py)

threshold_start = 1.0     # Start of decaying pseudo‑label threshold
threshold_decay = 0.9999  # Threshold decay
embed_dim       = 1024    # ImageBind huge text/video embedding size
gnn_hidden      = 8       # per‑node hidden dim inside KnowledgeGCN

Modify any field and re‑run train.py; everything else adapts automatically.

---

📝 Citation

@inproceedings{yun2025missiongnn,
  title={MissionGNN: Hierarchical Multimodal GNN-based Weakly Supervised Video Anomaly Recognition with Mission-Specific Knowledge Graph Generation},
  author={Yun, Sanggeon and Masukawa, Ryozo and Na, Minhyoung and Imani, Mohsen},
  booktitle={2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  pages={4736--4745},
  year={2025},
  organization={IEEE}
}