MultiADS: Defect-aware Supervision for Multi-type Anomaly Detection and Segmentation in Zero-Shot Learning

Paper

Purpose of the project

This software is a research prototype, solely developed for and published as part of the publication MultiADS: Defect-aware Supervision for Multi-type Anomaly Detection and Segmentation in Zero-Shot Learning.

Introduction

We present MultiADS, the first framework that goes beyond binary “good/bad” inspection to detect, localize, and label multiple defect types simultaneously. Here, we propose the first benchmark for the new task multi-type anomaly segmentation. Meanwhile, MultiADS also achieves state-of-the-art zero- and few-shot performance across six industrial datasets. Explore our code, pre-print, and live demo below!

Installation

Environment

1. Clone this repo

   git clone https://github.com/boschresearch/MultiADS.git
   cd MultiADS

2. Create & activate a conda environment

   conda env create -f environment.yml
   conda activate MultiADS
   

Datasets

Please download the datasets of MVTec-AD, VisA, MPDD, MAD, and Real-IAD into the data/ folder. Organize them as follows:

data
├── mvtec
│   └── meta.json
├── visa
│   ├── meta.json
│   └── meta_wo_md.json
├── mpdd
│   └── meta.json
├── MAD_Real
│   └── meta.json
├── MAD_Sim
│   └── meta.json
└── real_iad
    └── meta.json

To download our extend visa dataset with segmentation mask for each individual defects. Please visit VisA_Extended. Note that in the other datasets, each image contains only one type of defect types.

Training

python train.py \
  --dataset mvtec  \
  --train_data_path ./data/mvtec  \
  --save_path ./exps/mvtec_default/  

• dataset: dataset to use options = {mvtec, visa, mpdd, real_iad}
• train_data_path: where your training images reside
• save_path directory to write checkpoints & logs

Testing

Test Binary Anomaly Detection and Segmenation (Zero-shot)

python test.py \
  --dataset visa \
  --data_path ./data/visa \
  --save_path ./results/visa/zero_shot/ \
  --checkpoint_path ./exps/mvtec/epoch_1.pth 

Test Binary Anomaly Detection and Segmenation (K-shot)

python test.py \
  --dataset visa \
  --data_path ./data/visa \
  --save_path ./results/visa/few_shot/ \  
  --checkpoint_path ./exps/mvtec/epoch_1.pth \ 
  --k_shot 4 

• dataset: dataset to evaluate on
• data_path: path to your VisA test images
• save_path: where to write K-shot results
• checkpoint_path: which trained weights to load
• k_shot: few-shot number

Test Binary Anomaly Detection and Segmentation (Domain Adaption)

python test_da.py \
  --dataset visa   \
  --data_path ./data/visa \
  --save_path ./results/visa/domain_adption/  \
  --checkpoint_path ./exps/mvtec_default/epoch_1.pth 

Test Multi-type Anomaly Segmentation

python test_multi_defect.py 
    --dataset visa \               
    --data_path ./data/visa \
    --save_path ./results/mvtec_visa_multi_seg/zero_shot/ \
    --checkpoint_path ./exps/mvtec/epoch_1.pth \

Citation

@article{MultiADS2025,
  author       = {Ylli Sadikaj and
                  Hongkuan Zhou and
                  Lavdim Halilaj and
                  Stefan Schmid and
                  Steffen Staab and
                  Claudia Plant},
  title        = {MultiADS: Defect-aware Supervision for Multi-type Anomaly Detection
                  and Segmentation in Zero-Shot Learning},
  journal      = {CoRR},
  volume       = {abs/2504.06740},
  year         = {2025}
}