box_scores all equal to 0.5 and not available when using openVINO

[SebastienThibert]

Hello,

I previously used Anomalib v0.7 in detection mode, where I could retrieve a box score for each bounding box, making it easier to apply a threshold later.

However, in Anomalib v1.2, all bounding box scores are now set to 0.5:

Additionally, when exporting the model using OpenVINO, the box_scores are missing entirely:

Here are the details of my setup:

# Create the dataset using Anomalib's Folder dataset loader
datamodule = Folder(
    name="custom", 
    root=RAW_DATA_DIR,
    normal_dir="normal",
    abnormal_dir="abnormal",
    mask_dir="masks",
    image_size=INPUT_SIZE, 
    train_batch_size=BATCH_SIZE,
    eval_batch_size=BATCH_SIZE,
    task=TASK,
    num_workers=4,
    normal_split_ratio=0.2,  # Splitting 20% of normal data for validation
)

# Setup (splits train/val/test)
datamodule.setup()

# Load a batch from the validation set
val_loader = datamodule.val_dataloader()
i, data = next(enumerate(val_loader))

# Print available keys in the batch
print(data.keys())

# Initialize the Padim model
model = Padim(
    backbone=BACKBONE,
    layers=["layer1", "layer2", "layer3"],
)

engine = Engine(
    normalization=NormalizationMethod.MIN_MAX,
    threshold="F1AdaptiveThreshold",
    task=TASK,
    image_metrics=["AUROC"],
    accelerator="auto",
    check_val_every_n_epoch=1,
    devices=1,
    max_epochs=1,
    num_sanity_val_steps=0,
    val_check_interval=1.0,
)

engine.fit(model=model, datamodule=datamodule)

Please could you help me on this topic ?

[samet-akcay]

@adrianboguszewski, the issue you have recently experienced.. Is that similar to this?

[adrianboguszewski]

In my case classification score was always 0.