RIVAL10 Dataset

Classes

RIVAL10 adopts CIFAR10 classes via Imagenet. That is, two Imagenet synsets are combined for each RIVAL10 class, as visualized above.

Analysis

RIVAL10 has many uses, from guiding models to learn from attributes to assessing model sensitivities to various regions. The latter was the focus of RIVAL10's original paper. Namely, noise and saliency analyses were used to assess background sensitivity and find spurious features. Below, we show examples where models appear more sensitive to backgrounds.

1 / 7

2 / 7

3 / 7

4 / 7

5 / 7

6 / 7

7 / 7

❮ ❯

Examples of highly accurate models showing greater sensitivity to backgrounds than foregrounds.

1 / 4

2 / 4

3 / 4

4 / 4

❮ ❯

Spurious features discovered by finding misalignments of saliencies with RIVAL10 object masks.

Citation

Please cite our CVPR paper if RIVAL10 is of use to you.

@inproceedings{moayeri2022comprehensive,
    title     = {A Comprehensive Study of Image Classification Model Sensitivity to
                 Foregrounds, Backgrounds, and Visual Attributes},
    author    = {Moayeri, Mazda and Pope, Phillip and Balaji, Yogesh and Feizi, Soheil},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
    month     = {June},
    year      = {2022},
    }