Pytorch re-implementation of the paper Uninformed Students: Student-Teacher Anomaly Detection with Discriminative Latent Embeddings by Paul Bergmann et al.
Unsupervised anomaly segmentation is a desirable and challenging task in many domains of computer vision. Existing works mainly focus on generative algorithms such as Generative Adversarial Networks (GANs) or Variational Autoencoders (VAEs). These detect anomalies using per-pixel reconstruction errors or by evaluating the density obtained from the model’s probability distribution. This has been shown to be problematic due to inaccurate reconstructions or poorly calibrated likelihoods [1] [2].
In this paper, a novel framework is proposed for unsupervised anomaly segmentation based on student–teacher learning, and demonstrated the state-of-the-art performance on MNIST, CIFAR10 and MVTec AD dataset.
Unfortunately, the official implementation is not published. Therefore I tried to re-implement it from scratch.
My teacher network is the feature extraction module of torchvision.models.vgg19(pretrained=True). The architecture of a student network is exactly same to a teacher network, but pretrained=False. The training dataset consists of 6,742 images with 1 label, which are treated as the anomaly-free data, and all images with 0~9 labels are used as the test dataset. Since the image size is very small, the entire images are input to the student and teacher network. The teacher descriptors serve as surrogate labels, and the goal of training is to make the student network regress the surrogate labels. As you can see, the parameters in the teacher network are fixed, and the parameters in the student network are updated with the backpropagation. The implementation is available here
During inference, 0~9 images are input to the teacher and student, and the anomaly score is computed. I expect that 1 images result in the low anomaly scores, but other numbers are unseen during training, so the anomaly scores will be higher. As a result, it worked as I expected. It is relatively difficult to distinguish between 1 and 7, which are similar in shape.
Next, I evaluated on the much more challenging MVTec AD dataset. The architecture of teacher and student is exactly same to the above MNIST experiment. Training dataset consists of 209 images with good label, and all images with good, broken_small, broken_large or contamination labels are used as the test dataset. The size of the original images is 900×900 , and 128×128 randomly cropped images are input to the networks, and the parameters in the student network are updated to regress the teacher descriptors. The implementation is available here
To achieve pixel-precise anomaly detection, in other word, to get the anomaly map, the 595,984 forward computations are required. (900-128) × (900-128) = 595,984. Then the 595,984 anomaly scores are arranged in the anomaly map by associating them with the center row and column of the patch. It takes an hour per an anomaly map. Please note that the computation costs should be reduced by knowledge distillation and fast dense local feature extraction, but I didn't do anything.
Next, I’m showing the anomaly map from broken_small image. If the anomalies are included in the cropped image, the anomaly score will be higher (low → high, blue → yellow). As you can see in the image at the bottom right, the area with the high anomaly score and the anomaly mask is overlaid, and my anomaly segmentation model seems to work well.
- TensorRT
- nn.Unfold
- Fast dense feature extraction
- ProbabilisticCrop
- An ensemble of student networks
- Multiscale prediction
git clone git@github.com:TaikiInoue/STAD.git
cd STAD
make docker_build
make docker_run
make python_run