TFDS now supports the Croissant 🥐 format! Read the documentation to know more.

siscore

Description:

SI-Score (Synthetic Interventions on Scenes for Robustness Evaluation) is a dataset to evaluate robustness of image classification models to changes in object size, location and rotation angle.

In SI-SCORE, we take objects and backgrounds and systematically vary object size, location and rotation angle so we can study the effect of changing these factors on model performance. The image label space is the ImageNet label space (1k classes) for easy evaluation of models.

More information about the dataset can be found at https://github.com/google-research/si-score

Additional Documentation: Explore on Papers With Code
Homepage: https://github.com/google-research/si-score
Source code: tfds.datasets.siscore.Builder
Versions:
- 1.0.0 (default): Initial release.
Auto-cached (documentation): No
Feature structure:

FeaturesDict({
    'dataset_label': ClassLabel(shape=(), dtype=int64, num_classes=1000),
    'image': Image(shape=(None, None, 3), dtype=uint8),
    'image_id': int64,
    'label': ClassLabel(shape=(), dtype=int64, num_classes=1000),
})

Feature documentation:

Feature	Class	Shape	Dtype
	FeaturesDict
dataset_label	ClassLabel		int64
image	Image	(None, None, 3)	uint8
image_id	Tensor		int64
label	ClassLabel		int64

Supervised keys (See as_supervised doc): ('image', 'label')
Citation:

@misc{djolonga2020robustness,
      title={On Robustness and Transferability of Convolutional Neural Networks},
      author={Josip Djolonga and Jessica Yung and Michael Tschannen and Rob Romijnders and Lucas Beyer and Alexander Kolesnikov and Joan Puigcerver and Matthias Minderer and Alexander D'Amour and Dan Moldovan and Sylvain Gelly and Neil Houlsby and Xiaohua Zhai and Mario Lucic},
      year={2020},
      eprint={2007.08558},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}