FastFlow-main

An unofficial PyTorch implementation of FastFlow: Unsupervised Anomaly Detection and Localization via 2D Normalizing Flows (Jiawei Yi et al.)

This code is based on gathierry's FastFlow. I make some changes (e.g. save segmentation result images on evaluate phase). Thanks gathierry.

Installation

Install packages with:

pip install -r requirements.txt

Dataset

Download MVTec-AD dataset.

The MVTec-AD dataset is organized as follows:

dataset-path/bottle/train/good/                  # train (normal)
dataset-path/bottle/test/good/                   # test  (normal)
dataset-path/bottle/test/defect-type/            # test  (abnormal)
dataset-path/bottle/ground_truth/defect-type/    # GT    (abnormal)

Train

Use CaiT as example

python main.py -cfg configs/cait.yaml --data_path path/to/dataset -cat category

Evaluate

You can also show the result images in visualizations folder.

python main.py -cfg configs/cait.yaml --eval --data_path path/to/dataset -cat category -ckpt _fastflow_experiment_checkpoints/exp[index]/[epoch#]

Performance

Image-level AUROC

category	bottle	cable	capsule	carpet	grid	hazelnut	leather	metal_nut	pill	screw	tile	toothbrush	transistor	wood	zipper
ours	1.000	0.999	0.994	1.000	0.994	0.999	1.000	1.000	0.984	0.951	1.000	0.958	1.000	1.000	0.997
paper	1.000	1.000	1.000	1.000	0.997	1.000	1.000	1.000	0.994	0.978	1.000	0.944	0.998	1.000	0.995

Pixel-level AUROC

category	bottle	cable	capsule	carpet	grid	hazelnut	leather	metal_nut	pill	screw	tile	toothbrush	transistor	wood	zipper
ours	0.977	0.982	0.990	0.994	0.976	0.993	0.996	0.982	0.989	0.994	0.970	0.992	0.974	0.960	0.985
paper	0.977	0.982	0.991	0.994	0.983	0.991	0.995	0.985	0.982	0.994	0.963	0.989	0.973	0.970	0.987

Sample Results