Describe the bug

When training a Patchcore model and setting the image size with

pre_processor = Patchcore.configure_pre_processor(image_size=(768, 1536))

it automatically changes the resolution to a lower one than specified. This does not happen if you include center_crop, such as

pre_processor = Patchcore.configure_pre_processor(image_size=(768, 1536), center_crop_size=(768, 1536))

I tracked this down to the file:
https://github.com/openvinotoolkit/anomalib/blob/v2.0.0/src/anomalib/models/image/patchcore/lightning_model.py#L188

Specifically the code:

if center_crop_size is None:
    # scale center crop size proportional to image size
    height, width = image_size
    center_crop_size = (int(height * (224 / 256)), int(width * (224 / 256)))

under def configure_pre_processor.

Using my previous image_size of (768, 1536), it changes it to (672, 1344), which checks out math-wise. Setting the image_size to (672, 1344) further downscales it, etc. I also noticed that other models (Padim, STFPM, Reverse Distillation) do not have a configure_pre_processor function. They instead use anomalib_module.py. I assume this is intended since I am not very knowledgeable about Anomalib.

The heat in the heatmaps also become offset due to this. Furthermore, if you place an anomaly at the bottom of the image (or other sides I assume), Patchcore does not pick this up at all due to the offset. It does not reflect in the anomaly score and shows no heat in the heatmap. Setting a center_crop also fixes this. I would show images of this, but I am currently doing my bachelor's thesis and the dataset has been provided by a company that may not accept me publishing images from their dataset here.

On a sidenote, should I be using 224 x 224 images for Patchcore, or does 256 x 256 work fine? I noticed that if you don't set a resolution it automatically downscales to 224 x 224, but for other models it downscales to 256 x 256.

Dataset

Folder

Model

PatchCore

Steps to reproduce the behavior

Some code here is irrelevant. Tested with and without tiling & visualizer.

import torch
from anomalib.data import Folder
from anomalib.engine import Engine
from anomalib.models import Patchcore
from anomalib.callbacks import TilerConfigurationCallback
from anomalib.visualization import ImageVisualizer

tiler_config_callback = TilerConfigurationCallback(enable=True, tile_size=[256, 256], stride=64)
pre_processor = Patchcore.configure_pre_processor(image_size=(768, 1536)) # Height, Width

visualizer = ImageVisualizer(
    fields=["image", "anomaly_map"], # Only visualize image and anomaly map (no masks)
    overlay_fields=[("image", ["anomaly_map"])], # Only overlay anomaly map on image
    field_size=(1536, 768), # Set size (Width, Height)
    output_dir="results/Patchcore/visualizations", # Set output directory
)

datamodule = Folder(
    root="./datasets/dataset",
    name="dataset",
    normal_dir="ok",
    normal_test_dir="ok_test",
    abnormal_dir="nok",
    val_split_mode='same_as_test',
    train_batch_size=1,
    eval_batch_size=1,
    num_workers=11,
)

datamodule.setup()

model = Patchcore(
    backbone="wide_resnet50_2",
    layers=["layer2", "layer3"],
    pre_trained=True,
    coreset_sampling_ratio=0.1,
    num_neighbors=9,
    pre_processor=pre_processor,
    visualizer=visualizer
)

engine = Engine(callbacks=[tiler_config_callback])
engine.fit(model=model, datamodule=datamodule)

then use engine.test or engine.predict...

Image resolution can be seen with print(prediction.image.shape)

OS information

OS: Windows 11
Python version: 3.11.9
Anomalib version: 2.0
PyTorch version: 2.6.0
CUDA/cuDNN version: 12.4
GPU models and configuration: 1x GeForce RTX 3060 Ti
Any other relevant information: I use Jupyter Notebook

Expected behavior

The images would maintain the specified resolution.

Screenshots

No response

Pip/GitHub

pip

What version/branch did you use?

No response

Configuration YAML

N/A

Logs

N/A

Code of Conduct

• I agree to follow this project's Code of Conduct