UniAD

Official PyTorch Implementation of A Unified Model for Multi-class Anomaly Detection, Accepted by NeurIPS 2022 Spotlight.

1. Quick Start

1.1 MVTec-AD

• Create the MVTec-AD dataset directory. Download the MVTec-AD dataset from here. Unzip the file and move some to ./data/MVTec-AD/. The MVTec-AD dataset directory should be as follows.

|-- data
    |-- MVTec-AD
        |-- mvtec_anomaly_detection
        |-- json_vis_decoder
        |-- train.json
        |-- test.json

• cd the experiment directory by running cd ./experiments/MVTec-AD/.

• Train or eval by running:

  (1) For slurm group: sh train.sh #NUM_GPUS #PARTITION or sh eval.sh #NUM_GPUS #PARTITION.

  (2) For torch.distributed.launch: sh train_torch.sh #NUM_GPUS #GPU_IDS or sh eval_torch.sh #NUM_GPUS #GPU_IDS, e.g., train with GPUs 1,3,4,6 (4 GPUs in total): sh train_torch.sh 4 1,3,4,6.

  Note: During eval, please set config.saver.load_path to load the checkpoints.

• Results and checkpoints.

1.2 CIFAR-10

• Create the CIFAR-10 dataset directory. Download the CIFAR-10 dataset from here. Unzip the file and move some to ./data/CIFAR-10/. The CIFAR-10 dataset directory should be as follows.

|-- data
    |-- CIFAR-10
        |-- cifar-10-batches-py

• cd the experiment directory by running cd ./experiments/CIFAR-10/01234/. Here we take class 0,1,2,3,4 as normal samples, and other settings are similar.

• Train or eval by running:

  (1) For slurm group: sh train.sh #NUM_GPUS #PARTITION or sh eval.sh #NUM_GPUS #PARTITION.

  (2) For torch.distributed.launch: sh train_torch.sh #NUM_GPUS #GPU_IDS or sh eval_torch.sh #NUM_GPUS #GPU_IDS.

  Note: During eval, please set config.saver.load_path to load the checkpoints.

• Results and checkpoints. Training on 8 GPUs (NVIDIA Tesla V100 16GB) results in following performance.

2. Visualize Reconstructed Features

We highly recommend to visualize reconstructed features, since this could directly prove that our UniAD reconstructs anomalies to their corresponding normal samples.

2.1 Train Decoders for Visualization

• cd the experiment directory by running cd ./experiments/train_vis_decoder/.

• Train by running:

  (1) For slurm group: sh train.sh #NUM_GPUS #PARTITION.

  (2) For torch.distributed.launch: sh train_torch.sh #NUM_GPUS #GPU_IDS #CLASS_NAME.

  Note: for torch.distributed.launch, you should train one vis_decoder for a specific class for one time.

2.2 Visualize Reconstructed Features

• cd the experiment directory by running cd ./experiments/vis_recon/.

• Visualize by running (only support 1 GPU):

  (1) For slurm group: sh vis_recon.sh #PARTITION.

  (2) For torch.distributed.launch: sh vis_recon_torch.sh #CLASS_NAME.

  Note: for torch.distributed.launch, you should visualize a specific class for one time.

3. Questions

3.1 Explanation of Evaluation Results

The first line of the evaluation results are shown as follows.

The pixel means anomaly localization results.

The mean, max, and std mean post-processing methods for anomaly detection. That is to say, the anomaly localization result is an anomaly map with the shape of H x W. We need to convert this map to a scalar as the anomaly score for this whole image. For this convert, you have 3 options:

• use the mean value of the anomaly map.
• use the max value of the (averagely pooled) anomaly map.
• use the std value of the anomaly map.

In our paper, we use max for MVTec-AD and mean for CIFAR-10.

3.2 Visualize Learned Query Embedding

If you have finished the training of the main model and decoders (used for visualization) for MVTec-AD, you could also choose to visualize the learned query embedding in the main model.

• cd the experiment directory by running cd ./experiments/vis_query/.

• Visualize by running (only support 1 GPU):

  (1) For slurm group: sh vis_query.sh #PARTITION.

  (2) For torch.distributed.launch: sh vis_query_torch.sh #CLASS_NAME.

  Note: for torch.distributed.launch, you should visualize a specific class for one time.

Some results are very interesting. The learned query embedding partly contains some features of normal samples. However, we did not fully figure out this and this part was not included in our paper.

Acknowledgement

We use some codes from repositories including detr and efficientnet.