[Paper Review] Noise or Signal: The Role of Image Backgrounds in Object Recognition

Kai Xiao, Logan Engstrom+2 more|arXiv (Cornell University)|Jun 17, 2020

Domain Adaptation and Few-Shot Learning27 references59 citations

TL;DR

The paper analyzes how state-of-the-art object classifiers rely on image backgrounds, introducing a synthetic dataset toolkit to disentangle foreground and background signals and showing backgrounds can strongly influence predictions and robustness.

ABSTRACT

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.

Motivation & Objective

• Assess how much background signals contribute to object recognition accuracy on ImageNet-like data.
• Develop a dataset toolkit to disentangle foreground versus background information in images.
• Quantify models’ reliance on background signals and their robustness to background changes.
• Investigate how model accuracy and dependence on backgrounds have evolved with newer architectures and training regimes.

Proposed method

• Create ImageNet-9 (IN-9), a nine-class coarse-grained subset with foreground/background segmentation.
• Generate eight IN-9 variations to decouple foreground and background signals (e.g., Only-BG, No-FG, Mixed-Same, Mixed-Rand, Mixed-Next).
• Train and evaluate multiple architectures (e.g., ResNet-50, Wide-ResNet-50x2, MobileNet-v3, EfficientNet) on IN-9L and IN-9 variants.
• Define BG-Gap as the accuracy drop between Mixed-Same and Mixed-Rand to measure background reliance.
• Examine adversarial backgrounds and compute adversarial success rates per foreground class.
• Train models on Mixed-Rand to reduce reliance on background signals and assess robustness.

Experimental results

Research questions

• RQ1 To what extent do image backgrounds suffice for correct classification in modern models?
• RQ2 How do foreground and background signals contribute to model predictions across IN-9 variations?
• RQ3 Do more accurate models rely more or less on background cues, and can training regimes reduce background dependence?
• RQ4 How vulnerable are models to adversarially chosen backgrounds, and can background-robust training mitigate this?
• RQ5 How has background dependence evolved with ImageNet pre-training versus IN-9L training?

Key findings

• Backgrounds alone can enable non-trivial classification accuracy, well above random chance, on background-only datasets.
• A significant BG-Gap exists, with background signals causing misclassification when backgrounds are adversarial or mismatched to foregrounds.
• A substantial fraction of images (about 35%) require background information for correct classification when trained on Original datasets.
• More accurate models tend to exploit background correlations less or become more robust to background changes, though background signals persist, and pre-trained ImageNet models show reduced background dependence.
• Training on Mixed-Rand reduces reliance on backgrounds, making models more foreground-focused but sometimes reducing overall accuracy on Original data.
• Adversarial backgrounds can fool up to 87.5% of foreground examples, highlighting vulnerability in out-of-distribution settings.