celesnity-research/MMAD_remake_Benchmark

🏆 MMAD Remake — Unified Industrial Anomaly Detection Benchmark

Welcome to the MMAD Remake Benchmark Dataset, a comprehensive, unified Multimodal Visual Question Answering (VQA) and detailed image captioning benchmark built for evaluating Frontier Vision-Language Models (VLMs) across industrial quality control and visual anomaly detection.

---

📌 Dataset Overview

• Combined Datasets (4 Sources):
  1. GoodsAD: Supermarket goods anomaly detection (4,875 images)
  2. MVTec AD: Industrial surface & object anomaly detection (5,354 images)
  3. VisA: High-resolution printed circuit boards & multi-instance objects (10,821 images)
  4. MVTec LOCO AD: Structural and logical constraint anomalies (3,122 images)
• Total Images: 24,172 high-resolution images
• Total VQA Question Set: ~120,000+ multiple-choice questions across 7 core industrial subtasks.
• Caption Model: Qwen2-VL-2B-Instruct
• VQA Generator Model: Qwen2.5-7B-Instruct (4-bit Quantized)

---

📂 Directory Structure

MMAD_remake_Benchmark/
├── images/                             # Contains image files organized by source dataset
│   ├── goodsad/                        # 6 supermarket product categories
│   ├── mvtecad/                        # 15 industrial object & texture categories
│   ├── visa/                           # 12 PCB & multi-instance categories
│   └── mvtecloco/                      # 5 logical constraint categories
├── mmad_remake_captions.json           # Unified image captions across all 24,172 images
├── mmad_remake_questions.json          # Unified Multiple-Choice VQA Benchmark questions
├── mmad_remake_metadata.csv            # Overview metadata table for easy filtering
└── README.md                           # Benchmark documentation

---

📑 Core VQA Subtasks Evaluated

1. Anomaly Detection: Binary verification (Yes / No) if a defect exists.
2. Anomaly Classification: Identifying the exact defect category/type.
3. Anomaly Localization: Identifying the relative spatial location of the defect.
4. Attribute QA: Inquiring about physical attributes (colors, OCR text, branding, material).
5. Structural Reasoning: Evaluating component alignment, symmetry, and structural integrity.
6. Logical Reasoning: Checking for missing parts, extra items, or placement constraint violations.
7. Complex Decision Making: Final Quality Control pass/reject determination.

---

🛠️ Usage Example (Python PyTorch / HuggingFace)

import os
import json
import pandas as pd
from PIL import Image

BENCHMARK_ROOT = "./MMAD_remake_Benchmark"
QUESTIONS_FILE = os.path.join(BENCHMARK_ROOT, "mmad_remake_questions.json")

with open(QUESTIONS_FILE, "r", encoding="utf-8") as f:
    benchmark_data = json.load(f)

print(f"Total benchmark samples: {len(benchmark_data)}")

# Load a sample item
sample = benchmark_data[0]
image_full_path = os.path.join(BENCHMARK_ROOT, sample["image_path"])

# Open image
img = Image.open(image_full_path)
print(f"Sample Global ID: {sample['global_id']} (Source: {sample['dataset_source']})")
print(f"Loaded image: {image_full_path} (Size: {img.size})")

# Iterate questions
for q in sample["questions"]:
    print(f"\n[{q['subtask']}] {q['question']}")
    for opt in q["options"]:
        print(f"  {opt}")
    print(f"Correct Answer: {q['answer']}")

---

📜 License & Citation

The original images belong to their respective dataset creators (PKU-GoodsAD, MVTec Software GmbH, Amazon Research). Please cite the original papers when conducting non-commercial research using this benchmark:

@inproceedings{mmad_remake_2026,
  title={MMAD Remake: A Unified Multimodal Benchmark for Industrial Visual Anomaly Detection},
  author={Celesnity Research},
  year={2026}
}