LACToS: Local Anomaly Detection via Conditional Typicality of Superpixels
Local anomaly detection in images with multimodal inlier distributions—such as medical scans, faces, and X-ray baggage—remains challenging for industrial-defect methods that rely on heuristic reconstruction errors or feature distances. LACToS introduces a principled alternative by scoring superpixels via statistical typicality under the conditional inpainting posterior of an inlier diffusion model, made tractable by grouping pixels into superpixels and using a single denoising posterior as a cheap proxy. This shift from point estimates to distribution-aware testing matters because it directly addresses the false positives that arise when normal images contain diverse, context-dependent content.
LACToS offers a conceptually attractive and theoretically grounded approach to local out-of-distribution detection. The move from pixel-wise reconstruction error to superpixel-level conditional typicality is well-motivated, and the denoising proxy is both clever and supported by an explicit approximation bound. Empirically, the method is competitive across medical, facial, and security benchmarks and exhibits lower false-positive rates under controlled multimodality. Nevertheless, the small test-set sizes, lack of statistical testing, and a hyperparameter selection protocol that appears to use the test set itself temper the strength of the empirical claims. Overall, it is a solid contribution with notable limitations.
The strongest aspect of the paper is its clear theoretical motivation: the authors identify that collapsing a posterior over reconstructions into a first-order statistic destroys distributional information and propose typicality as a remedy. The two corresponding engineering innovations are both practical and backed by analysis. The toy experiment in Section 3 isolates multimodality as a failure mode for PatchCore and DDAD and shows LACToS remains robust. Theorem D.1 in Appendix D rigorously bounds the error introduced by replacing per-superpixel inpainting posteriors with a single denoising posterior, lending credibility to the computational shortcut.
The paper relies heavily on Assumption 4.1, which posits that pixels within a superpixel are conditionally i.i.d. given their complement. This is implausible for natural images where spatial correlation within superpixels is high, and the method’s histogram fitting step depends directly on it; the authors concede the assumption may be violated.
More troublingly, Appendix E.3 states that hyperparameters were chosen by grid search over a subset of 20 test images to maximize AUROC before evaluating on the full test set. If that subset was drawn from the reported test set, this constitutes data leakage and inflates the metrics. Even if it was a separate validation split, its small size and the large search space raise doubt about generalization. Finally, the real-world evaluations use extremely small test sets (50–100 images per condition) and no significance tests, making head-to-head comparisons with baselines statistically unreliable.
The toy experiment in Section 3 provides direct, controlled evidence that FPR rises with inlier multimodality for DDAD and PatchCore while LACToS controls it, exactly matching the paper’s central claim. On real benchmarks, however, the evidence is less decisive. The authors themselves note that “the best method differs across metrics” (Table 1). LACToS leads in AP and AUPRO on CHAOS and FFHQ, but PatchCore slightly exceeds it in AUROC on DVXray, and DDAD surpasses it in AUROC on CHAOS. The baselines are appropriate and retrained per dataset, yet the comparison omits other distribution-aware diffusion baselines that might also benefit from typicality. The use of synthetic tumors in CHAOS and a manually curated, tiny OOD face set further limits ecological validity.
The authors pledge to release code and provide extensive hyperparameter tables in Appendix E, which aids reproduction. However, independent reproduction faces serious hurdles: training an unconditional diffusion model per dataset is expensive, and LACToS depends on many interrelated hyperparameters including the DPS noise level $\sigma_{\mathrm{noise}}$, superpixel average size $S_{\mathrm{avg}}$, histogram bins $B_{\mathrm{RGB}}$ and $B_{\mathrm{feat}}$, histogram smoothing $\sigma_{\mathrm{hist}}$, feature weight $\lambda$, and sample count $M$. The reported selection of these parameters via AUROC-maximizing grid search on 20 test images is the single biggest reproducibility concern because it blurs the line between validation and test data. Additionally, the computational cost of diffusion posterior sampling is orders of magnitude higher than feedforward methods like PatchCore, raising practical barriers to adoption.
Local anomaly detection is the task of identifying spatially confined anomalous or out-of-distribution (OOD) regions within an otherwise in-distribution (ID) image. Most existing methods are designed for a specific context, with a strong focus on industrial defect detection, where they achieve excellent performance on well-established benchmarks. However, we observe that these methods can exhibit significant performance drops in other contexts, especially when the inlier distribution is multimodal. To account for this inlier complexity, we propose LACToS, a principled local OOD detection method based on statistical typicality and made tractable through two innovations: (i) grouping pixels into superpixels and (ii) using a computationally efficient proxy to estimate typicality. We evaluate LACToS across three diverse applications spanning medical imaging, human faces, and X-ray security screening. LACToS performs more robustly than prior methods across these benchmarks. The code is released here: https://github.com/voilalab/LACToS.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.