FiLo: Zero-Shot Anomaly Detection by Fine-Grained Description and High-Quality Localization

Official implementation of paper FiLo: Zero-Shot Anomaly Detection by Fine-Grained Description and High-Quality Localization (ACM MM 2024).

Introduction

Welcome to the official repository for "FiLo: Zero-Shot Anomaly Detection by Fine-Grained Description and High-Quality Localization." This work presents FiLo, an innovative method for Zero-Shot Anomaly Detection (ZSAD) that addresses the challenges of detecting and localizing anomalies without prior knowledge of normal or abnormal samples.

FiLo comprises two key components: Fine-Grained Description (FG-Des) and High-Quality Localization (HQ-Loc). FG-Des leverages Large Language Models (LLMs) to generate detailed anomaly descriptions for each object category, enhancing the accuracy and interpretability of anomaly detection. HQ-Loc improves localization by combining preliminary localization using Grounding DINO, position-enhanced text prompts, and a Multi-scale Multi-shape Cross-modal Interaction (MMCI) module, allowing for precise anomaly detection across various sizes and shapes.

Overview of FiLo

Running FiLo

Environment Installation

Clone the repository locally:

git clone https://github.com/CASIA-IVA-Lab/FiLo.git

Install the required packages:

pip install -r requirements.txt

Prepare Grounding DINO checkpoint

You can download our fine-tuned Grounding DINO model from the table below. We fine-tuned Grounding DINO using MMDetection. Consistent with FiLo's experimental setup, we tested Grounding DINO fine-tuned on the VisA dataset on the MVTec dataset and tested Grounding DINO fine-tuned on the MVTec dataset on the VisA dataset.

Training dataset	Grounding DINO Weights Address
MVTec	groundingdino_train_on_mvtec
VisA	groundingdino_train_on_visa

Prepare FiLo checkpoint

You can download our pre-trained FiLo checkpoint from the table below.

Training dataset	FiLo Weights Address
MVTec	filo_train_on_mvtec
VisA	filo_train_on_visa

Prepare data

MVTec AD

• Download and extract MVTec AD into data/mvtec
• runpython data/mvtec.py to obtain data/mvtec/meta.json

data
├── mvtec
    ├── meta.json
    ├── bottle
        ├── train
            ├── good
                ├── 000.png
        ├── test
            ├── good
                ├── 000.png
            ├── anomaly1
                ├── 000.png
        ├── ground_truth
            ├── anomaly1
                ├── 000.png

VisA

• Download and extract VisA into data/visa
• runpython data/visa.py to obtain data/visa/meta.json

data
├── visa
    ├── meta.json
    ├── candle
        ├── Data
            ├── Images
                ├── Anomaly
                    ├── 000.JPG
                ├── Normal
                    ├── 0000.JPG
            ├── Masks
                ├── Anomaly
                    ├── 000.png

Test our model

You can refer to the parameter settings in test.sh to modify the dataset path and checkpoint path for testing.

bash test.sh

Train your own weights

bash train.sh

Citation:

If you found FiLo useful in your research or applications, please kindly cite using the following BibTeX:

@article{gu2024filo,
  title={FiLo: Zero-Shot Anomaly Detection by Fine-Grained Description and High-Quality Localization},
  author={Gu, Zhaopeng and Zhu, Bingke and Zhu, Guibo and Chen, Yingying and Li, Hao and Tang, Ming and Wang, Jinqiao},
  journal={arXiv preprint arXiv:2404.13671},
  year={2024}
}