cs.CVNov 16, 2026

Signs of Wear: Evaluating Intelligent Speed Assistance Under Synthetic Adverse Conditions

Authors: Johannes Fabian Hahn, Mao Shan, Julie Stephany Berrio Perez

Organizations: Department of Informatics, University of Hamburg, Germany. · Intelligent Transport Systems group at the Australian Center for Robotics, The University of Sydney, Australia. · School of Mechanical and Mining Engineering, University of Queensland, Australia.

Abstract

Intelligent Speed Assistance (ISA) systems rely on cameras to recognize speed-limit signs, yet existing benchmarks underrepresent the weathered, dirty, and overgrown signs commonly encountered on rural roads. We use generative image editing to introduce different forms of wear into real street-level images of speed-limit signs. The resulting ISA-Wear dataset contains paired original and edited images. Guided by text instructions, a multimodal image-editing model adds snow, rust, vegetation, faded paint, or mud to the speed-limit sign in a street-level frame at three severity levels. Where appropriate, the surrounding scene is adapted as well, and the viewpoint is kept. Every edited frame has an unedited counterpart, and we only evaluate frames whose original is recognized correctly, so any failure can be attributed to the wear. Since the model alters the sign's geometry in 21% of the edits, all edits are filtered by a geometric check and a manual review. Severity is measured as the LPIPS distance between the original and the edited sign plate. On 716 worn frames, we find that in detect-then-read pipelines, wear breaks detection before it breaks reading. The detector finds only 62% of the worn signs, even though it finds every one of them in the original frames. Missed detections account for 78--87% of failed reads, a pattern that a zero-shot detector shares. Given the annotated sign box, as in the crop-based benchmarks, a classifier still reads 78% of the worn signs, but only 56% when it has to rely on the detector

Figures & tables

Explore similar work

Jul 31, 2026

Explaining AI-Image Detection: What the Heatmap Actually Shows

A marketplace review photograph is a document: platforms approve refunds on it, and generative models drove the cost of forging one to zero. We study that detection problem, so we build a detector and attach an attribution map as its evidence, then measure what that pair delivers on 186,527 images under controls designed to change our conclusions when something is wrong. Compression history, not synthesis, drives naive evaluation: our strongest model reaches 0.9999 PR-AUC (area under the precision-recall curve) on a product-disjoint split, yet falls to 0.7254 once we re-encode synthetics into the real class's format, while five public detectors move by at most 0.07. Aligning one class relocates the cue rather than removing it, and the repaired model then assigns native files a median probability of synthesis of 0.0004. One identical final encode for both classes repairs that, and a three-seed factorial credits the encoding change with the whole gain (+0.176 +- 0.009 PR-AUC). That encode equalises the last stage only: forensic features alone still separate the classes at 0.7145 against a base rate of 0.254. For evidence we test maps causally, against controls that never consult the detector. Whether an attribution ranking exists at all depends on whether the detector reacts to the image. On our first-fix detector, which calls 96 of 100 edited frames real, no map beats a random one. On the detector we selected, twelve of seventeen maps clear that control on edited images and eight on generated ones; perturbation leads both axes and no gradient-CAM variant shows a positive advantage. The trivial controls never clear it, and on generated images the centre prior is worse than random. Our ensembled regional map clears both axes and takes the top pixel AP at 12.4 s per map against 44.9 for occlusion. Clearing a detector-blind control is not yet a faithful explanation, and we demonstrate none.
Jul 16, 2026

GlobalForge: Towards Robust AI-Generated Image Detection

AI-generated image (AIGI) detectors achieve strong accuracy on clean benchmarks, but their performance drops sharply after images are propagated through real-world channels. We trace this fragility to what these detectors actually learn: they overfit to local artifacts left by generators in small spatial neighborhoods, which are easily destroyed by common propagation degradations such as JPEG compression and blur. Instead, we shift the discriminative cue from fragile local artifacts to more robust global structure. Building on this, we propose GlobalForge, a framework with two complementary modules. The Local Information Bottleneck (LIB) suppresses local components to block shortcut learning, while the Global Structural Reasoning (GSR) module forces every token to gather evidence from distant regions. Both modules are trained jointly under a contrastive structural loss based on degradation that keeps the resulting features stable under degradation. To support fine-grained robustness evaluation, we further introduce RealDeg-Bench, covering 7 common degradation operators and multi-step compound chains. GlobalForge improves average BAcc on 8 in-the-wild benchmark groups by 5.89%\mathbf{5.89\%} over the previous state-of-the-art, and is clearly ahead of representative baselines on RealDeg-Bench under both single and compound degradations. Code is available at https://anonymous.4open.science/r/GlobalForge-BE0F/.
Jan 30, 2026

AI-Generated Image Detectors Exploit Global Artifacts: Evidence from Inpainting Exchange

Modern diffusion-based inpainting enables realistic, localized image manipulation, raising critical challenges for reliable detection. We show that state-of-the-art AI-generated image detectors largely fail to identify what was synthesized and instead exploit a global artifact introduced as a side effect of the generation pipeline: latent-diffusion inpainting passes the entire image through a VAE encoder--decoder, inducing a subtle but pervasive spectral shift even in unedited regions. We turn this observation into a simple, model-free diagnostic, Inpainting Exchange (INP-X), that restores the original pixels outside the edited region while preserving all synthesized content. To study the effect at scale, we build a benchmark of 3030K matched real, standard-inpainted, and exchanged triplets (9090K images) spanning four datasets and three inpainters. INP-X requires no model access and is a single compositing step, yet it collapses pretrained detectors, commercial APIs, and bias-aware methods toward chance (commercial APIs: >$$90\%\!\to\!\sim$$55\%; the bias-aware DDA: 97.6% ⁣→ ⁣58.4%97.6\%\!\to\!58.4\%). We give formal propositions linking the effect to high-frequency attenuation from the VAE information bottleneck, and we causally isolate it with a factorial battery and a variance decomposition that attributes 97%97\% of the background spectral shift to the autoencoder round-trip alone. A perceptual study (n=153n{=}153) shows the shift is also perceptible to humans, whose accuracy falls from 63.8%63.8\% to 56.2%56.2\% (p=0.002p{=}0.002): significantly reduced, though not eliminated. Finally, we show that mechanism-driven mitigations (mixed-distribution and localization-supervised training) improve robustness, but detecting INP-X edits remains an open problem. Our benchmark and code are available at: https://github.com/emirhanbilgic/INP-X