2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2022
DOI: 10.1109/cvpr52688.2022.01850
|View full text |Cite
|
Sign up to set email alerts
|

A Comprehensive Study of Image Classification Model Sensitivity to Foregrounds, Backgrounds, and Visual Attributes

Help me understand this report

Search citation statements

Order By: Relevance

Paper Sections

Select...
2
2
1

Citation Types

0
9
0

Year Published

2022
2022
2024
2024

Publication Types

Select...
4
1
1

Relationship

1
5

Authors

Journals

citations
Cited by 24 publications
(9 citation statements)
references
References 21 publications
0
9
0
Order By: Relevance
“…Spurious feature reliance becomes problematic under distribution shifts that break their correlation to class labels: sidewalk segmentation struggles in the absence of cars [56], familiar objects cannot be recognized in unfamiliar poses [1] or uncommon settings [27,5], etc. A natural and ubiquitous spurious correlation in vision is image backgrounds, observed in numerous prior works to be leveraged by models for classification [68,41] and object detection [47]. Spurious correlations also relate to algorithmic biases [13,17], with implications for fairness [19,7,9,28], reflecting the importance of this issue.…”
Section: Review Of Literaturementioning
confidence: 99%
See 4 more Smart Citations
“…Spurious feature reliance becomes problematic under distribution shifts that break their correlation to class labels: sidewalk segmentation struggles in the absence of cars [56], familiar objects cannot be recognized in unfamiliar poses [1] or uncommon settings [27,5], etc. A natural and ubiquitous spurious correlation in vision is image backgrounds, observed in numerous prior works to be leveraged by models for classification [68,41] and object detection [47]. Spurious correlations also relate to algorithmic biases [13,17], with implications for fairness [19,7,9,28], reflecting the importance of this issue.…”
Section: Review Of Literaturementioning
confidence: 99%
“…Importantly, this effect does not hold for distribution shifts that maintain spurious correlations, indicating that the reduced distributional robustness is due to increased spurious feature reliance. We now directly quantify sensitivity to core features via RIVAL10 and Salient ImageNet-1M datasets [41,60]. The premise of this analysis is that model sensitivity to an input region can be quantified by the drop in accuracy due to corrupting that region [59].…”
Section: At Hurts Natural Distributional Robustness Only When Spuriou...mentioning
confidence: 99%
See 3 more Smart Citations