Weird Anomaly scores

[rishabh-akridata]

Hello,
Trained the model on the imagenet30 dataset and inference it on a few images. However, the anomaly scores are coming in some weird numbers. Something like 84.31, 84.32, etc for both anomalous and nonanomalous images. The script that I am using for inference is attached below. Is there something wrong I am doing?

from uniformaly.uniformaly import Uniformaly
device = "cuda"
backbone_name = "dino_vit_base_8"
layers_to_extract_from = ["2", "3", "4", "5", "6", "7", "8", "9"]
params = {"backbone": backbone_name, "layers_to_extract_from": layers_to_extract_from}
model = Uniformaly(device, params)

model_path = "dino_vit_base_8/models/imagenet_acorn"
model.load_from_path(load_path=model_path, device="cuda", local_nn_method=uniformaly.common.FaissNN(False, 4), topk=0.05, thres=0.1)

test_images = [os.path.join("imagenet30_subset/one_class_test/acorn/n12267677", i) 
for i in os.listdir("imagenet30_subset/one_class_test/acorn/n12267677")]

from datasets.transforms import ImageTransform
from tqdm import tqdm
resize=256 
imagesize=224
transform_img = ImageTransform(resize, imagesize)
prediction_scores = []
for img in tqdm(test_images, total=len(test_images)):
  preprocessed_image = transform_img(Image.open(img))
  predictions = model.predict(preprocessed_image.unsqueeze(0))
  anomaly_score = predictions[0][0]
  prediction_scores.append(anomaly_score)

Thanks

[sy00n]

Hello, thank you for your interest in our paper.

How many images were used for evaluation during the test? Also, are the 84.31 and 84.32 anomaly scores or AUROC values?

[rishabh-akridata]

@sy00n These are anomaly scores, that I am getting when I am inferencing the model on a few images.

[sy00n]

I cannot give a definitive answer as I cannot verify which image it is, but the anomaly images might be confusing examples. Since the anomaly score is a distance measure, only such small differences may occur.

If you have any other issues or questions, please let us know. Thank you