Datasets:
license: cc-by-nc-4.0
dataset_info:
features:
- name: image
dtype: image
- name: world_name
dtype: string
- name: character_name
dtype: string
- name: character_label
dtype: string
- name: character_rotation_yaw
dtype: int64
- name: character_rotation_roll
dtype: int64
- name: character_rotation_pitch
dtype: int64
- name: character_scale
dtype: float64
- name: camera_roll
dtype: int64
- name: camera_pitch
dtype: int64
- name: camera_yaw
dtype: int64
- name: character_texture
dtype: string
- name: scene_light
dtype: string
splits:
- name: train
num_bytes: 29382707151.112
num_examples: 88328
download_size: 29358745565
dataset_size: 29382707151.112
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
PUG: ImageNet
The PUG: ImageNet dataset contains 88,328 pre-rendered images based on Unreal Engine using 724 assets representing 151 ImageNet classes with 64 environments, 7 sizes, 9 textures, 18 different camera orientations, 18 different character orientations and 7 light intensities. In contrast to PUG: Animals, PUG: ImageNet was created by varying only a single factor at a time (which explains the lower number of images than PUG: Animals despite using more factors). The main purpose of this dataset is to provide a novel, useful benchmark, paralleling ImageNet, but for fine-grained evaluation of the robustness of image classifiers, along several factors of variation.
LICENSE
The datasets are distributed under the CC-BY-NC, with the addenda that they should not be used to train Generative AI models.
Citing PUG
If you use one of the PUG datasets, please cite:
@misc{bordes2023pug,
title={PUG: Photorealistic and Semantically Controllable Synthetic Data for Representation Learning},
author={Florian Bordes and Shashank Shekhar and Mark Ibrahim and Diane Bouchacourt and Pascal Vincent and Ari S. Morcos},
year={2023},
eprint={2308.03977},
archivePrefix={arXiv},
primaryClass={cs.CV}
}