LACToS: Local Anomaly Detection via Conditional Typicality of Superpixels

Jihoon Hong; Alireza Kheirandish; Rohan Sanda; Sara Fridovich-Keil

cs.CVOct 22, 2026

LACToS: Local Anomaly Detection via Conditional Typicality of Superpixels

Authors: Jihoon Hong, Alireza Kheirandish, Rohan Sanda, Sara Fridovich-Keil

Organizations: Georgia Institute of Technology, Atlanta, GA, USA · Stanford University, Stanford, CA, USA · Cohere

Abstract

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.

Figures & tables

Figure 1: Using a diffusion model pretrained on ID images only, LACToS is able to identify locally OOD regions in images across various tasks, including tumors in CT scans, anomalies such as scars on faces, and harmful objects such as scissors in X-ray baggage scans.

Figure 2: Multimodality Toy Experiment. (Top) Images from the inlier distributions across different complexity levels. The top-right corner always contains a circle, while other corners can contain different shapes, with a wider variety of shapes at higher levels. (Bottom) False positive rates at true positive rate of 95 $\%$ , evaluated on images that contain anomalous squares in the top-right corner.

Figure 3: A fine-grained superpixel decomposition of an example image.

Dataset	Metric	PatchCore	SimpleNet	SuperSimpleNet	EfficientAD	GLASS	DDAD	LACToS (Ours)
CHAOS	AUROC	0.9297	0.7464	0.3786	0.6664	0.8318	0.9707	0.9620
	AP	0.0590	0.0193	0.0051	0.0105	0.0410	0.3159	0.4318
	AUPRO	0.6811	0.4191	0.0332	0.2513	0.5530	0.8676	0.8764
FFHQ	AUROC	0.8491	0.7162	0.7647	0.6200	0.6707	0.7469	0.8515
	AP	0.0798	0.0537	0.0772	0.0405	0.0345	0.0563	0.1177
	AUPRO	0.4094	0.3174	0.4047	0.2365	0.2440	0.3083	0.5320

Table 1: Pooled quantitative performance in local anomaly detection. AUROC and AP are computed over all pixels in each dataset jointly. AUPRO is computed over the dataset using per-region overlap for

\mathrm{FPR}\leq 0.3

. Higher is better for all metrics. The best result is shown in bold and the second-best is underlined .

Figure 5: Anomaly heatmaps generated by each method on example images from each test dataset. Rows 1-2 ( CHAOS ): detect liver tumors in CT scans. Rows 3-4 ( FFHQ ): detect anomalous facial features. Rows 5-6 ( DVXray ): detect weapons (scissors) in X-ray baggage scans.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Level	Set of Candidate Shapes	Probability
Level 1	$\{\text{Circle, Square}\}$	$[0.5,0.5]$
Level 2	$\{\text{Circle, Square, Triangle}\}$	$[0.\overline{3},0.\overline{3},0.\overline{3}]$
Level 3	$\{\text{Circle, Square, Triangle, Ellipse}\}$	$[0.25,0.25,0.25,0.25]$
Level 4	$\{\text{Circle, Square, Triangle, Ellipse, Star}\}$	$[0.2,0.2,0.2,0.2,0.2]$

Appendix

Table 2: Toy Experiment 1 Inlier Distribution Specification.

Figure 6: Example Images from Training Sets. The shape orientations are random, with positions sampled with jitter around fixed anchors in each quadrant. The upper-right corner, marked with a red box, always contains a circle. With increasing level, each of the other three corners has increasing modality, as more shapes are equally probable under the inlier distribution.

Figure 7: LACToS Toy Experiment 1: Robustness Under Multimodality. Average per-image False Positive Rates at True Positive Rate $\in\{80\%,90\%,95\%,99\%\}$ , evaluated on images that contain squares on the upper-right corner.

Level	Set of Candidate Shapes	Probability
Level 1	$\{\text{Circle, Square}\}$	$\{0.5,0.5\}$
Level 2	$\{\text{Circle, Square}\}$	$\{0.6,0.4\}$
Level 3	$\{\text{Circle, Square}\}$	$\{0.7,0.3\}$
Level 4	$\{\text{Circle, Square}\}$	$\{0.8,0.2\}$
Level 5	$\{\text{Circle, Square}\}$	$\{0.9,0.1\}$

Appendix

Table 3: Toy Experiment 2 Inlier Distribution Specification.

Figure 8: Example Images from Training Sets. Positions of the shapes are sampled with jitter around fixed anchors, with random orientations. The upper-right corner, marked with a red box, always contains a circle. With increasing level, the probability of a square at the other three corners decreases, making it more difficult to distinguish squares there from the square in the upper-right corner.

Figure 9: LACToS Toy Experiment 2. AUROCs across different inlier square probabilities, over methods and different sample sizes. (Left) AUROC is computed over pixels that belong to inlier squares and circles, with those pixels from squares labeled positive, and negative otherwise. Higher AUROC means the method is more likely to distinguish an inlier square (minority) from inlier circles (majority). (Right) AUROC is computed over pixels that belong to squares, with those pixels from the square in the upper-right corner (truly OOD) labeled positive, and negative otherwise (minority inlier). Higher AUROC means the method better distinguishes an inlier but uncommon square from a strictly out-of-distribution square.

Dataset	# train (inlier)	# test (ID + OOD)	Eval mask	Resolution
CHAOS	3,070	100 + 100	body silhouette	$256{\times}256$
FFHQ	5,000	50 + 50	—	$256{\times}256$
DVXray (scissors)	5,000	72 + 72	—	$600{\times}800$
MVTec AD (15 categories)	3,629	1,258	—	$256{\times}256$

Appendix

Table 4: Dataset composition. Training images contain only inlier (anomaly-free) examples. Test images contain local anomalies and are paired with pixel-level ground-truth masks.

Parameter	CHAOS	FFHQ	DVXray
DPS $\sigma_{\mathrm{noise}}$	0.1	0.01	2.0
DPS scale	0.4	0.1	1.0
$S_{\mathrm{avg}}$	100	20	10
$M$	5	20	20
$\sigma_{\mathrm{hist}}$	0.5	1.0	0
$(B_{\mathrm{RGB}},B_{\mathrm{feat}})$	$(16,64)$	$(16,32)$	$(16,16)$

Appendix

Table 5: Dataset-specific hyperparameters used for LACToS.

Parameter	CHAOS	FFHQ	DVXray
PatchCore	$k=1$	$k=1$	$k=1$
EfficientAD	$\alpha=0$	$\alpha=0.7$	$\alpha=0.2$
DDAD	$(w,v)=(2,1)$	$(w,v)=(2,4)$	$(w,v)=(4,4)$

Appendix

Table 6: Dataset-specific hyperparameters used for baseline methods.

Figure 10: CHAOS Additional Visualizations. 10 randomly selected test images and the heatmaps for each method.

Figure 11: FFHQ Additional Visualizations. 10 randomly selected test images and the heatmaps for each method.

Figure 12: DVXray Additional Visualizations. 10 randomly selected test images and the heatmaps for each method.

Method	Metric	Carpet	Grid	Leather	Tile	Wood	Bottle	Cable	Capsule
SimpleNet	AUROC	0.9765	0.9656	0.9898	0.9326	0.9177	0.9775	0.9688	0.9862
	AP	0.3650	0.2440	0.3257	0.5470	0.3673	0.6315	0.6151	0.3729
	AUPRO	0.9047	0.8757	0.9630	0.7781	0.7866	0.8945	0.8741	0.9093
SuperSimpleNet	AUROC	0.9832	0.9771	0.9878	0.9383	0.9208	0.9779	0.9728	0.9805
	AP	0.4548	0.2383	0.3492	0.5109	0.3867	0.6202	0.6151	0.3973
	AUPRO	0.9222	0.9236	0.9686	0.8107	0.8551	0.8982	0.8746	0.8753

Appendix

Table 7: Results on MVTec AD. We report pixel-level AUROC, AP, and AUPRO for each object or texture category. AUROC and AP pool every pixel of all test images of the category (anomalous and anomaly-free); AUPRO is the region-based per-region overlap integrated up to an FPR of 0.3. Each method occupies three rows, one per metric. Higher is better for all metrics.

Dataset	Metric	$M=1$		$M=5$		$M=10$		$M=20$
Dataset	Metric	DDAD	LACToS	DDAD	LACToS	DDAD	LACToS	DDAD	LACToS
CHAOS	AUROC	0.9707	0.9552	0.9748	0.9620	0.9749	0.9615	0.9745	0.9620
	AP	0.3159	0.2536	0.3496	0.4318	0.3468	0.4236	0.3489	0.4338
	AUPRO	0.8676	0.8524	0.8827	0.8764	0.8825	0.8713	0.8814	0.8729
FFHQ	AUROC	0.7469	0.8446	0.7512	0.8521	0.7559	0.8503	0.7583	0.8515
	AP	0.0563	0.0937	0.0543	0.1197	0.0552	0.1119	0.0564	0.1177

Appendix

Table 8: Ablation on sample size

M

in diffusion-based methods. AUROC, AP, and AUPRO are reported for each dataset.

Dataset	Metric	Average size of superpixels
Dataset	Metric	$10$	$20$	$30$	$50$	$70$	$100$	$200$
CHAOS	AUROC	$0.9418$	$0.9504$	$0.9535$	$0.9578$	$0.9595$	$0.9620$	$0.9593$
	AP	$0.1271$	$0.1599$	$0.1857$	$0.2434$	$0.2643$	$0.4318$	$0.3677$
	AUPRO	$0.7992$	$0.8274$	$0.8385$	$0.8499$	$0.8571$	$0.8764$	$0.8600$
FFHQ	AUROC	$0.7930$	$0.8515$	$0.8502$	$0.8480$	$0.8462$	$0.8438$	$0.8383$
	AP	$0.0774$	$0.1177$	$0.1145$	$0.1107$	$0.1095$	$0.1065$	$0.0998$

Appendix

Table 9: Ablation on the average size of superpixels. The larger the average size, the smaller the total number of superpixels. AUROC, AP, and AUPRO on CHAOS , FFHQ , and DVXray .

Dataset	Metric	$\lambda=0$	$\lambda=0.25$	$\lambda=0.5$	$\lambda=0.75$	$\lambda=1.0$	$\lambda=1.5$	$\lambda=2$	$\lambda=4$	Feature only
CHAOS	AUROC	0.8418	0.9638	0.9632	0.9625	0.9620	0.9613	0.9609	0.9601	0.9592
	AP	0.0729	0.3893	0.4318	0.4356	0.4318	0.4192	0.4056	0.3645	0.2733
	AUPRO	0.5689	0.8764	0.8778	0.8772	0.8764	0.8755	0.8748	0.8735	0.8717
FFHQ	AUROC	0.7605	0.8515	0.8488	0.8459	0.8437	0.8406	0.8386	0.8347	0.8293
	AP	0.0677	0.1177	0.1115	0.1075	0.1041	0.0997	0.0975	0.0944	0.0916
	AUPRO	0.3688	0.5320	0.5353	0.5345	0.5331	0.5305	0.5287	0.5252	0.5211

Appendix

Table 10: Ablation on the feature-weighting parameter

\lambda

\lambda=0

corresponds to using RGB features only. AUROC, AP, and AUPRO are reported on CHAOS , FFHQ , and DVXray .

Superpixel Segmentation

See all 9 papers

Statistical Non-linear Reconstruction Loss for Image Anomaly Detection
Jul 14, 2026Nguyen Minh Tri, Hoang Khuong Duy, Huynh Cong Viet NguIndustrial Anomaly Detection Anomaly Localization
AIC-Lab, FPT University
Reasoning-Driven Anomaly Detection and Localization with Image-Level Supervision
Mar 28, 2026Yizhou Jin, Yuezhu Feng, Jinjin Zhang +3Anomaly Localization Multimodal Anomaly Detection
State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, Beijing, China · Hangzhou Innovation Institute, Beihang University, Hangzhou, China
NFAD: Nuisance-Filtered Anomaly Detection Under Distribution Shift
Aug 29, 2026Dat Cao, Son Nghiem, Phan Nguyen +2Anomaly Detection Industrial Anomaly Detection
Korea Advanced Institute of Science and Technology (KAIST), South Korea · Sony Computer Science Laboratories (Sony CSL), Japan · The University of Tokyo, Japan +1

Jul 14, 2026J/K move · Enter open · S save

Statistical Non-linear Reconstruction Loss for Image Anomaly Detection

Nguyen Minh Tri, Hoang Khuong Duy, Huynh Cong Viet Ngu

AIC-Lab, FPT University

Industrial Anomaly Detection Anomaly Localization Reconstruction-Based Anomaly Detection Unsupervised Anomaly Detection

Open paper AskPDF arXiv

Reconstruction-based methods are a cornerstone of unsupervised image anomaly detection, but they remain vulnerable to outlier leakage, where standard mean squared error (MSE) loss drives the model to faithfully reconstruct anomalous patterns. We propose a Non-linear Reconstruction Loss that applies a sigmoid-based squashing function to suppress high-magnitude features, preventing outliers from dominating optimization while preserving sensitivity to normal patterns. In addition, we introduce a statistical calibration scheme that selects the scaling factor

k

from the confidence interval (CI) of the normal feature distribution, enabling data-driven control of the suppression strength. Our approach achieves competitive or superior anomaly detection performance compared to state-of-the-art methods, reaching 99.0% Image-AUROC and 97.3% Pixel-AUROC on MVTec-AD, and 95.3% Image-AUROC and 99.0% Pixel-AUROC on VisA. These results indicate that non-linear gradient suppression is an effective mechanism for mitigating outlier leakage and improving anomaly localization in unified industrial inspection settings. The implementation is available at https://github.com/mintii13/Statistical-Non-linear-Reconstruction-Loss.git.

Abstract

Superpixel Segmentation

Explore similar work

Statistical Non-linear Reconstruction Loss for Image Anomaly Detection

Reasoning-Driven Anomaly Detection and Localization with Image-Level Supervision

NFAD: Nuisance-Filtered Anomaly Detection Under Distribution Shift