Research map / Detection and segmentation / Open-set and few-shot detection
Noise or Signal: The Role of Image Backgrounds in Object Recognition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train on Mixed-Rand so background labels are decorrelated from foreground labels | ResNet-50 · ImageNet-9 Mixed-Rand | accuracy: +17.3% | Helped |
| Train on Mixed-Rand so background labels are decorrelated from foreground labels | ResNet-50 · ImageNet-9 Mixed-Next | accuracy: +22.3% | Helped |
| Train on Mixed-Rand so background labels are decorrelated from foreground labels | ResNet-50 · ImageNet-9 Only-BG-T | accuracy: 15% accuracy | Hurt |
| Use class-consistent backgrounds rather than randomized backgrounds at test time | ImageNet-pretrained models · ImageNet-9 Mixed-Same vs Mixed-Rand | accuracy: 13-22% and 6-14% | Helped |
| Replace the background with an adversarially selected background | standard object classifiers · ImageNet-9 adversarial backgrounds | misclassification rate as background class: 87.5% | Hurt |
| Train on more fine-grained classes (ablation) | IN-9LB classifiers · ImageNet-9 Mixed-Same vs Mixed-Rand | BG-gap: 2.3% smaller | Helped |
| Train on more fine-grained classes (ablation) | IN-9LB classifiers · ImageNet-9 Original | accuracy | Hurt |
| Increase training dataset size (ablation) | IN-9LB classifiers · ImageNet-9 Mixed-Same vs Mixed-Rand | BG-gap | No clear effect |
| Increase training dataset size (ablation) | IN-9LB classifiers · ImageNet-9 test datasets | accuracy | Helped |
| Use ℓp-robust training (ablation) | ResNet-50 · ImageNet-9 Mixed-Same vs Mixed-Rand | BG-gap | No clear effect |
| Train on ImageNet rather than IN-9L | ResNet-50 · ImageNet-9 Mixed-Same vs Mixed-Rand | BG-gap: 14.3% -> 6.0% | Helped |
| Provide a background consistent with the foreground class | Original-trained ResNet-50 · ImageNet-9 Original | correct classification requiring background: 35% | Mixed |
Closest papers by meaning
- ObjectNet: A large-scale bias-controlled dataset for pushing the limits of object recognition models (NeurIPS 2019 · similarity 0.90)
- A Comprehensive Study of Image Classification Model Sensitivity to Foregrounds, Backgrounds, and Visual Attributes (CVPR 2022 · similarity 0.88)
- Spatially Invariant Unsupervised Object Detection with Convolutional Neural Networks (AAAI 2019 · similarity 0.87)
- Contemplating Real-World Object Classification (ICLR 2021 · similarity 0.87)
- ContextShift: A Controlled Benchmark for Context Dependence in Object Detection (NeurIPS 2026 · similarity 0.86)
- Object Detection Benchmarks are Incomplete: The Role of Label Errors and Annotation Uncertainty (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.