Datasets:
metadata
license: cc-by-sa-4.0
arxiv: 2403.09193
dataset_info:
features:
- name: image
dtype: image
- name: lf_label
dtype: string
- name: hf_label
dtype: string
splits:
- name: test
num_bytes: 43555505.52
num_examples: 1280
download_size: 43548336
dataset_size: 43555505.52
configs:
- config_name: default
data_files:
- split: test
path: data/test-*
task_categories:
- image-classification
language:
- en
tags:
- imagenet
- frequncy-bias
- texture-shape-bias
- vision-language-model
pretty_name: Low-/High-Frequency Cue-Conflict
This dataset contains images of natural objects with low- and high-frequency cue conflicts, as described in the paper Can We Talk Models Into Seeing the World Differently?.
The dataset consists of 1200 images where the low-frequency and high-frequency information correlate with different classes (out of 16 classes). This dataset can be used to evaluate the frequency-bias of vision models.