Noise or Signal: The Role of Image Backgrounds in Object Recognition

By Kai Y. Xiao, Logan Engstrom, Andrew Ilyas, Aleksander Madry. 2020

466 citations recorded by the source corpus.

We assess the tendency of state-of-the-art object recognition models to depend on signals from image backgrounds. We create a toolkit for disentangling foreground and background signal on ImageNet images, and find that (a) models can achieve non-trivial accuracy by relying on the background alone, (b) models often misclassify images even in the presence of correctly classified foregrounds--up to 87.5% of the time with adversarially chosen backgrounds, and (c) more accurate models tend to depend on backgrounds less. Our analysis of backgrounds brings us closer to understanding which correlations machine learning models use, and how they determine models' out of distribution performance.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

Citations in the corpus

Projected GANs Converge Faster (2021, methodology): “This hypothesis follows from the fact that we used image classification models for the projection, which have been shown to only marginally reduce in accuracy when applied on images of objects with removed background [66].”
SWAD: Domain Generalization by Seeking Flat Minima (2021, methodology): “Table 6 show the generalizability of SWAD on ImageNet [42] and its shifted benchmarks, namely, ImageNet-C [59], ImageNet-R [60], and background challenge (BGC) [61].”
A Wholistic View of Continual Learning with Deep Neural Networks: Forgotten Lessons and the Bridge to Active and Open World Learning (2020, background): “…the data distribution, the obstacle of predicting overconﬁdently on unseen data is further magniﬁed by the ubiquitous option for any classiﬁer to diﬀerentiate a concept based on a combination of noise patterns, the absence of a speciﬁc pattern, or background patterns altogether (Xiao et al., 2020).”
WILDS: A Benchmark of in-the-Wild Distribution Shifts (2020, neutral): “…et al., 2018), ImageNet-C (Hendrycks and Dietterich, 2019), and similar ImageNet variants (Geirhos et al., 2018); and datasets that crop out objects and replace their backgrounds, as in the Backgrounds Challenge (Xiao et al., 2020) and other similar datasets (Sagawa et al., 2020; Koh et al., 2020).”
Last Layer Re-Training is Sufficient for Robustness to Spurious Correlations (2022, background): “[100] proposed several datasets in the Backgrounds Challenge to study the effect of the backgrounds on predictions of ImageNet models.”
Vision Transformers are Robust Learners (2021, methodology): “Through a set of carefully designed experiments, we ﬁrst verify the enhanced robustness of ViTs to common robustness benchmark datasets [13, 14, 9, 15].”
Rethinking Spatial Dimensions of Vision Transformers (2021, methodology): “Standard Occ IN-A [15] BGC [40] FGSM [10]”
Selective Classification Can Magnify Disparities Across Groups (2020, background): “, 2019), image backgrounds (Ribeiro et al., 2016; Xiao et al., 2020), spurious clinical variables (Badgeley et al.”
Delving Deep into the Generalization of Vision Transformers under Distribution Shifts (2021, unknown): “ImageNet9 [36] is adopted for background shifts.”
Counterfactual Generative Networks (2021, background): “Xiao et al. (2020) propose the BG-Gap to measure a classifier’s dependence on the background signal.”
Long-Short Transformer: Efficient Transformers for Language and Vision (2021, unknown): “In addition to out-of-distribution robustness (ImageNet-Real and Imageet-v2), we further investigate the robustness of our vision transformer against common corruption (ImageNet-C), semantic shifts (ImageNet-R), Background dependence (ImageNet-9) and natural adversarial examples (ImageNet-A).”
Re-labeling ImageNet: from Single to Multi-Labels, from Global to Localized Labels (2021, methodology): “Impact of ReLabel on FGSM [11], ImageNet-A [17], ImageNet-C [16], and background challenge [52] benchmarks.”
iBOT: Image BERT Pre-Training with Online Tokenizer (2021, methodology): “We evaluate this property on ImageNet-9 (IN-9) dataset (Xiao et al., 2020).”
Distilling Model Failures as Directions in Latent Space (2022, unknown): “More broadly, the training dataset might contain spurious correlations, encouraging the model to depend on prediction rules that do not generalize to deployment (Xiao et al., 2020; Geirhos et al., 2020; DeGrave et al., 2021).”