💡 What makes OOD evaluation hard? Real-world anomalies are too rare to collect at scale.
💡 Our answer: A Synthetic Anomaly Integration Pipeline that inserts physically-plausible OOD objects into existing 3D occupancy data — producing 3 benchmark datasets spanning 26 anomaly categories.

---

 🧪 Pipeline   📦 Datasets   🤖 Models   🚀 Quick Start 

---

🧪 Pipeline

Problem	OOD object data is scarce — real anomalies are rare and annotation is expensive
Solution	Synthetic anomaly generation under physical & environmental constraints
Result	Plausible, challenging OOD evaluation data at scale

---

📦 Datasets & Downloads

📥 3 benchmarks built on SemanticKITTI & KITTI-360 (click to collapse)

#	Dataset	Base	Anomaly Objects	Download
1	VAA-KITTI	SemanticKITTI	26 categories	⬇️ Google Drive
2	VAA-KITTI-360	KITTI-360	26 categories	⬇️ Google Drive
3	VAA-STU	STU	—	⬇️ Google Drive

📊 Anomaly types: animals · furniture · garbage bags · construction debris · vegetation overgrowth · vehicles · and more

---

🤖 Model Zoo

Dataset	Model	mIoU	Log	Weight
SemanticKITTI	OccOoD-T	16.80	log	pth
SemanticKITTI	OccOoD-S	13.79	log	pth
KITTI-360	OccOoD-T	18.38	log	pth
KITTI-360	OccOoD-S	12.47	log	pth

---

🚀 Quick Start

# 1. Install environment
#    → docs/install.md

# 2. Download datasets
#    → docs/dataset.md

# 3. Train & evaluate
#    → docs/run.md

📖 install.md  ·  dataset.md  ·  run.md

---