VID-AD: A Dataset for Image-Level Logical Anomaly Detection under Vision-Induced Distraction

Abstract

Logical anomaly detection in industrial inspection remains challenging due to variations in visual appearance (e.g., background clutter, illumination shift, and blur), which often distract vision-centric detectors from identifying rule-level violations. However, existing benchmarks rarely provide controlled settings where logical states are fixed while such nuisance factors vary. To address this gap, we introduce VID-AD, a dataset for logical anomaly detection under vision-induced distraction, containing 10 manufacturing scenarios and five capture conditions, which yields 50 one-class tasks and 10,395 images. Each scenario is defined by two logical constraints (from quantity, length, type, placement, and relation) and includes both single-constraint and combined violations. We further propose a language-based anomaly detection framework that relies solely on text descriptions generated from normal images. Using contrastive learning with positive texts and contradiction-based negative texts synthesized from these descriptions, our method learns embeddings that emphasize logical content rather than low-level appearance. Extensive experiments demonstrate consistent improvements over baselines across the evaluated settings.

Overview

Method

Our language-based anomaly detection framework leverages Vision Language Models (VLMs) to generate textual descriptions of normal images. We employ contrastive learning with BERT, using positive sentences from normal images and contradiction-based negative sentences to learn embeddings that emphasize logical content over low-level appearance variations.

Dataset

Download

The dataset can be downloaded from here.

Dataset Structure

VID-AD_dataset/
├── {Category}/                    # e.g., Balls, Blocks, Cookies, ...
│   ├── train/
│   │   └── good/                  # Normal training video frames
│   └── test/
│       ├── good/                  # Normal test video frames
│       └── logical_anomalies/     # Anomalous test video frames
│           ├── Single-Aspect-A/
│           ├── Single-Aspect-B/         # (varies by category)
│           └── Dual-Aspects/
├── {Category}_Cable_BG/           # Cable Background
├── {Category}_Mesh_BG/            # Mesh Background
├── {Category}_Blurry_CD/          # Blurry Condition
└── {Category}_Low-light_CD/       # Low-light Condition

10 Scenarios: Balls, Blocks, Cookies, Dishes, Fruits, Ropes, Stationery, Sticks, Tapes, Tools

Five Capture Conditions: Original, Cable_BG, Mesh_BG, Low-light_CD, Blurry_CD

Dataset Statistics

Results

Our method demonstrates consistent improvements over existing baseline methods across all capture conditions and categories.

Overall Performance

ROC-AUC scores across different capture conditions (White BG, Cable BG, Mesh BG, Low-light CD, Blurry CD) show that our approach maintains robustness against vision-induced distraction.

Per-Category Results

Detailed per-category (Stick, Fruits, Tools, Cookies, Tapes, Stationery, Ropes, Blocks, Dishes, Balls) error rates validate the effectiveness across diverse manufacturing scenarios.

Ablation Study

Ablation studies demonstrate the importance of each component in our framework.

Requirements

conda env create -f environment.yml
conda activate vid-ad_venv

Model Weights

The following pre-trained models are automatically downloaded from Hugging Face when first used. Access to some models may require accepting the license on the Hugging Face model page.

Model	Hugging Face ID	Note
Qwen2-VL (default)	Qwen/Qwen2-VL-7B-Instruct
Llama 3.2 Vision	meta-llama/Llama-3.2-11B-Vision-Instruct	Requires access approval
LLaVA v1.5	llava-hf/llava-1.5-13b-hf
BERT	google-bert/bert-base-uncased	Used for contrastive learning

To download models in advance, you can use the Hugging Face CLI:

pip install huggingface_hub
huggingface-cli login  # Required for gated models (e.g., Llama)
huggingface-cli download Qwen/Qwen2-VL-7B-Instruct  # Default model

For other models:

huggingface-cli download meta-llama/Llama-3.2-11B-Vision-Instruct  # Requires login
huggingface-cli download llava-hf/llava-1.5-13b-hf

Usage

Basic Usage

python verification.py --model qwen

Arguments

Argument	Description	Default
--model	VLM model to use (qwen, llama, or llava)	(required)
--datasets	Dataset(s) to process	All datasets
--conditions	Condition(s) to process (Original, Cable_BG, Mesh_BG, Low-light_CD, Blurry_CD)	All conditions
--base_dir	Base directory for output (results and models)	./output
--dataset_dir	Base directory of the VID-AD dataset	./dataset/VID-AD_dataset
--prompt_dir	Directory containing prompt files	./prompt

Examples

Process specific datasets and conditions:

python verification.py --model llama --datasets Balls Sticks --conditions Original Cable_BG

Specify custom paths:

python verification.py --model qwen \
    --base_dir /path/to/output \
    --dataset_dir /path/to/VID-AD_dataset \
    --prompt_dir /path/to/prompt

Prompt Files

The --prompt_dir directory should contain:

• {Category}_prompt.txt — Image description prompt for each category (e.g., Balls_prompt.txt)
• negative_sentence_prompt.txt — Prompt for generating negative (anomalous) sentences

License

This dataset is released under the MIT License.