tencentyouturesearch / anomalydetection_real-iad

AnomalyDetection_Real-IAD

This repo has been transferred to https://github.com/Tencent/AnomalyDetection_Real-IAD

anomalydetection_real-iad's Issues

dataset structure setting

Hello, I would like to ask about the structure of the Real-IAD dataset, what specific images correspond to the 'OK' and 'NG' folders under each category? What do the 'AK', 'HS', 'PS', and other folders in the 'NG' folder represent? Why are these classifications made?

Questions within the paper

Hello, thank you for proposing such an excellent dataset.

I have two questions to ask you:
The first one is Figure 2(b)(c), which explains the wider defect area and proportion. I don’t understand this part very well. What do the horizontal and vertical axes of (b) and (c) represent? And what do these two charts want to express?

The second question is about the evaluation indicators. I particularly want to know the specific algorithm of S-AUROC and whether there are program codes for these three indicators (S-AUROC/I-AUROC/P-PRO) that can help people understand the specific operations.

Thank you!

About “Real-IAD”

Hi, is the “Real-IAD” dataset open source yet? If so, where can I download it?

data processing

Hello, thank you for your excellent work. I would like to ask you how to adjust the data to 256 size mentioned in your paper. Do you directly resize the 1024 size data to 256? If so, will it lead to a huge loss of information?

How to integrate multiview results?

Thanks for your excellent work! Based on the description of the paper, the results from multiviews are merged to assess sample-level performance. Is it possible to tell us how to merge these results, especially when the anomaly is misdirected or detected incorrectly?

Having trouble to download the data from huggingface

Hi !

I just tried to download Real-IAD.

But an error occurred constantly..

the error is..

  File "/home/jisu/anaconda3/lib/python3.9/site-packages/huggingface_hub/hf_file_system.py", line 638, in get_file
    http_get(
  File "/home/jisu/anaconda3/lib/python3.9/site-packages/huggingface_hub/file_download.py", line 570, in http_get
    raise EnvironmentError(
OSError: Consistency check failed: file should be of size 23215841838 but has size 10001900996 ((…)f5878d/realiad_raw/rolled_strip_base.zip).
We are sorry for the inconvenience. Please retry with `force_download=True`.
If the issue persists, please let us know by opening an issue on https://github.com/huggingface/huggingface_hub.

The code I used is 

from datasets import load_dataset
# Real-IAD/Real-IAD 데이터셋 로드 및 다운로드
dataset = load_dataset("Real-IAD/Real-IAD", cache_dir="/SSD2/Datasets/RealIAD")

I downloaded the data above amount.

Resolving data files: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 123/123 [00:00<00:00, 477.45it/s]
Downloading data:  77%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████▏                                | 93/121 [1:51:55<5:05:29, 654.63s/files]
Downloading data:  87%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████▎                  | 105/121 [7:30:11<1:08:36, 257.25s/files]

Any solution??

thanks!

The single-view Real-IAD datasets

Thank you for your great work. In the article you mentioned using single-/multi-view Real-IAD to compare effects with other data sets such as Mvtec, I want to know how to get single-view Real-IAD, Do I need to extract files one by one？I need to get a structure similar to the Mvtec dataset for testing.

questions about UIAD setting

Hi, great work!
I was wondering, in the UIAD setting which Table 2 used, the model are trained on all categories in the dataset and tested on each category (the setting UniAD proposed)? or the model are trained on one specific category and tested on the corresponding category (one model for one category)?
Thanks in advance!

test result

Hello, I have an additional question. In your original article, you used the official code to get the results of DeSTSeg on your data. I used his code and there was no problem when training. When testing, if there is a lot of data, the memory will continue to increase until it overflows. Have you encountered this problem?