CRACKS Dataset

Abstract

Crowdsourcing annotations has created a paradigm shift in the availability of labeled data for machine learning. Availability of large datasets has accelerated progress in common knowledge applications involving visual and language data. However, specialized applications that require expert labels lag in data availability. One such application is fault segmentation in subsurface imaging. Detecting, tracking, and analyzing faults has broad societal implications in predicting fluid flows, earthquakes, and storing excess atmospheric CO2. However, delineating faults with current practices is a labor-intensive activity that requires precise analysis of subsurface imaging data by geophysicists. In this paper, we propose the CRACKS dataset to detect and segment faults in subsurface images by utilizing crowdsourced resources. We leverage Amazon Mechanical Turk to obtain fault delineations from sections of the Netherlands North Sea subsurface images from (i) 26 novices who have no exposure to subsurface data and were shown a video describing and labeling faults, (ii) 8 practitioners who have previously interacted and worked on subsurface data, (iii) one geophysicist to label 7636 faults in the region. Note that all novices, practitioners, and the expert segment faults on the same subsurface volume with disagreements between and among the novices and practitioners. Additionally, each fault annotation is equipped with the confidence level of the annotator. The paper provides benchmarks on detecting and segmenting the expert labels, given the novice and practitioner labels. Additional details along with the dataset links and codes are available at https://alregib.ece.gatech.edu/cracks-crowdsourcing-resources-for-analysis-and-categorization-of-key-subsurface-faults/.

Dataset Labels

The dataset comes provided with 400 images of seismic sections. This can be found under the column Image. For each Image, labels for them were generated by 26 novices, 8 practitioners, and an expert. The columns are named accordingly: expert for the expert, noviceX for novice number X, and practitionerX for practiotioner number X. Each label is composed of 3 three which correspond to the confidence of existing fault:

• Orange: No fault (Certain)
• Green: Falut (Uncertain)
• Blue: Fault (Certain)

However, not all seismic images have all the corresponding labels as some are missing the expert label as well as other labels. These correspond to panda's NaN and are viewed in HuggingFace's dataset viewer as Not supported with pagintation yet.

Installation and Usage

Provided is a sample code that downloads the dataset from HuggingFace and accesses the first Image and novice01 label. It then loads it into a pytorch DataLoader:

from datasets import load_dataset
from torch.utils.data import DataLoader

# Load the dataset from huggingface
cracks_train = load_dataset('gOLIVES/CRACKS', split='train')
cracks_test_1 = load_dataset('gOLIVES/CRACKS', split='test_1')
cracks_test_2 = load_dataset('gOLIVES/CRACKS', split='test_2')

# Accesses the first image and its first novice label
seismic_image = cracks_train[0]['Image']
novice_laebl = cracks_train[0]['novice01']

# Covert into a Format Usable by Pytorch
cracks_train = cracks_train.with_format("torch")

dataloader = DataLoader(cracks_train, batch_size=4)
for batch in dataloader:
    print(batch) 

Links

Associated Website: https://alregib.ece.gatech.edu/