ImageNet-LC: A Benchmark for Object-Centric Robustness Under Localized Corruptions

Computer Vision systems exhibit significant performance degradation under localized visual distortions, such as lens flare, occlusion, or sensor artifacts, that selectively affect semantic objects in real-world scenes. Existing robustness benchmarks, including ImageNet-C, predominantly evaluate global corruptions and therefore under-represent object-centric degradation regimes. To address this gap, we introduce ImageNet-LC (Localized Corruptions), a benchmark designed to evaluate robustness under object-centric, localized corruptions applied to regions of interest identified by an object detection model. ImageNet-LC comprises seven representative corruption types applied at multiple severity levels, enabling controlled analysis of object-level robustness. We evaluate a diverse set of CNNs and vision transformers models on ImageNet-LC and compare their performance against established benchmarks.