RobustAD Dataset

About the Dataset

RobustAD, specifically designed to evaluate the robustness of anomaly detection models in real-world scenarios. RobustAD features a curated dataset of defect detection images with meticulously controlled distribution shifts across multiple dimensions relevant to practical applications and more closely mirrors real-world deployment scenarios. RobustAD is designed to cover inspection challenges across multiple industries to ensure the diversity of use cases and encourage the development of generalizable methods. It is carefully curated to reflect the complexity of real-world anomaly detection task in terms of both the defect variations and the domain shifts captured in the data. Robus- tAD consists of 3 sub-datasets corresponding to 3 different objects of interest, each with a source domain data for training and multiple target domains with different shifts for testing. The PCB sub-dataset captures the challenges of finding subtle scratches, soldering melts, and missing parts which comprise of the most common defects encoun- tered during inspection of Printed Circuit Boards in electronics and semiconductor manufacturing. The metal parts sub-dataset reflects the challenges of inspecting metal automotive parts with reflective surfaces for possible chipping, dents, or porosity (holes in metal) in the automotive industry. The pile of packets represents a common count-based anomaly detection task performed by packaging machines in the pharmaceutical industry. We believe this broad cov- erage of tasks and anomaly types across important sectors ensures a general model that is relevant for common in- dustry inspection problems and serves as a good starting point. The PCB and metal parts datasets are defined for localization and classification tasks where as piled packets subset is only defined for classification task.

Dataset Card for RobustAD

For more details, refer to this paper: COMING SOON!

How to Use

To load the dataset,

from datasets import load_dataset
from datasets import Image

#For piled bags dataset (Classification only)
piled_bags_dataset = load_dataset("imagefolder", data_files={"train": 'PiledBags/piled_bags_data_dir_train/*', "test0": 'PiledBags/piled_bags_data_dir_test0/*' , "test1": 'PiledBags/piled_bags_data_dir_test1/*' , "test2": 'PiledBags/piled_bags_data_dir_test2/*' , "test3": 'PiledBags/piled_bags_data_dir_test3/*' ,"test4": 'PiledBags/piled_bags_data_dir_test4/*', "test5": 'PiledBags/piled_bags_data_dir_test5/*'})

#For PCB dataset
pcb_dataset = load_dataset("imagefolder", data_files={"train": 'PCB/pcb_data_dir_train/*', "test0": 'PCB/pcb_data_dir_test0/*', "test1": 'PCB/pcb_data_dir_test1/*' , "test2": 'PCB/pcb_data_dir_test2/*' , "test3": 'PCB/pcb_data_dir_test3/*' ,"test4": 'PCB/pcb_data_dir_test4/*', "test5": 'PCB/pcb_data_dir_test5/*'}).cast_column("mask", Image(decode=True))

#For Metal Parts dataset
metal_parts_dataset = load_dataset("imagefolder", data_files={"train": 'MetalParts/metal_parts_data_dir_train/*', "test0": 'MetalParts/metal_parts_data_dir_test0/*' , "test1": 'MetalParts/metal_parts_data_dir_test1/*' , "test2": 'MetalParts/metal_parts_data_dir_test2/*' , "test3": 'MetalParts/metal_parts_data_dir_test3/*' ,"test4": 'MetalParts/metal_parts_data_dir_test4/*', "test5": 'MetalParts/metal_parts_data_dir_test5/*', "test6": 'MetalParts/metal_parts_data_dir_test6/*'}).cast_column("mask", Image(decode=True))

#metal_parts_dataset['train'][0] - Normal sample does not have a mask
#{'image': <PIL.Image.Image image mode=RGB size=2681x1500 at 0x7F66A1BE46D0>, 'label': 0, 'mask': None}

#metal_parts_dataset['train'][0]  -  Anomaly samples have a mask
{'image': <PIL.Image.Image image mode=RGB size=2681x1500 at 0x7F66A1B1EBC0>, 'label': 1, 'mask': <PIL.PngImagePlugin.PngImageFile image mode=L size=2681x1500 at 0x7F66A1BE7040>}

License Information

The RobustAD dataset is released under the Creative Commons license cc-by-4.0.

Citation Information

COMING SOON!

Contact

[email protected] (Latha Pemula) | [email protected] (Dongqing Zhang) | [email protected] (Onkar Dabeer)