What does EfficientAD differ from PatchCore?

The features of EfficientAD

Determining normal (local) features through learning by using quite simple convolutional neural network
- Computation time in the inference process is quite fast (It is just simple convolution)
- The receptive field is NOT broad(33 by 33), and it is controllable
- Extracting features using distillation from ImageNet task
- Training process might take time
Considering logical (it probably means global) features by AutoEncoder
Compute anomaly score from the above two sets of local features/global features

Determining normal (local) features by sampling feature map from ResNet
- Training process for neural network is not necessary
- The receptive field highly depends on ResNet
- Nearest neighbor search is required for inference step
Not considering logical (global) features
Determine an anomaly score by comparing it with the sampled normal features mentioned above.

The network structure is very simple and has low computational complexity.
Learning normal features using Back Propagation eliminates the need for sampling normal features.
Measures are taken to avoid overfitting to normal features.
- Quantile Sampling is used to prevent overfitting to excessive normal features.
- A Penalty Term for the Student Network helps avoid ImageNet Bias.
The use of distillation makes it easier to control the Receptive Field (unlike PatchCore, which relies on ResNet/WideResNet).
The receptive field of EffectiveAD is easily controllable. It is exatcly within 33 pixels on initial setting. PatchCore exhibits a broader dependence, which may potentially have a negative impact on anomaly localization.
It can also consider global features in anomaly detection.