MoViAD: Modular Visual Anomaly Detection

• MoViAD is a modular visual anomaly detection framework that integrates ready-to-use models, datasets, and utilities for comprehensive VAD benchmarking.
• It supports diverse methods—including memory-bank, student–teacher, and reconstruction approaches—enabling robust performance in unsupervised, noisy environments.
• Designed for scalable deployment, MoViAD facilitates edge-to-cloud experimentation, continual learning, and extensive profiling for industrial and research applications.

MoViAD denotes “Modular Visual Anomaly Detection,” a modular deep learning library for Visual Anomaly Detection (VAD) that provides ready-to-use models, trainers, datasets, backbones, quantization and compression utilities, and evaluation and profiling infrastructure (Barusco et al., 16 Jul 2025). In closely related literature, near-homonymous names designate adjacent problems in online video anomaly detection, industrial multi-view anomaly understanding, and multiple visual artifact detection, while some works use “MoViAD”-style language for broader multi-modal or motion-video settings (Rossi et al., 2023, Zhao et al., 11 May 2026, Feng et al., 2024).

1. Nomenclature and scope

The name “MoViAD” is not fully uniform across recent preprints. The explicit title “MoViAD: Modular Visual Anomaly Detection” identifies a library for image-based VAD, whereas related works use MOVAD, MMVIAD, and MVAD for distinct but neighboring formulations. Some papers further describe “MoViAD”-type models only by analogy rather than as official system names. This suggests that the term has both a precise library-level meaning and a broader contextual use across anomaly-centric vision systems (Barusco et al., 16 Jul 2025, Rossi et al., 2023, Zhao et al., 11 May 2026, Feng et al., 2024).

Term	Expansion	Domain
MoViAD	Modular Visual Anomaly Detection	Visual anomaly detection library
MOVAD	Memory-augmented Online Video Anomaly Detection	Dashcam video anomaly detection
MMVIAD	Multi-view Multi-task Video Industrial Anomaly Detection	Industrial anomaly dataset and benchmark
MVAD	Multiple Visual Artifact Detector	Video streaming artifact detection

Two adjacent usages are notable. First, the CMViM paper states that “MoViAD” is not explicitly mentioned, but that CMViM is “essentially a prototype of what a ‘Multi-modal Vision Mamba for Alzheimer’s Disease’ (MoViAD) system could look like” (Yang et al., 2024). Second, the MoVi dataset paper presents “MoVi-based Action Detection / Analysis” as a natural interpretation of “MoViAD,” again without introducing the acronym as an official system name (Ghorbani et al., 2020).

2. MoViAD as “Modular Visual Anomaly Detection”

As a library, MoViAD is a modular deep learning framework for VAD whose design goal is fast and easy access to state-of-the-art VAD models and utilities, while also supporting practical deployment and extensibility (Barusco et al., 16 Jul 2025). It is organized around separate but interoperable modules for datasets, VAD methods, trainers, backbones, quantization and compression, evaluation, and other VAD utilities. This separation is central: the dataset abstraction handles preprocessing and splits, the backbone abstraction exposes intermediate feature maps, the model abstraction implements anomaly scoring, and the trainer abstraction manages method-specific optimization.

The implemented method families span the major contemporary VAD paradigms. Memory-bank methods include PatchCore, CFA, and PaDiM; student–teacher methods include STFPM, PaSTe, and RD4AD; reconstruction methods include DRAEM and Ganomaly; flow-based modeling is represented by FastFlow; and adversarial modeling includes SuperSimpleNet (Barusco et al., 16 Jul 2025). The library also integrates multiple public datasets—MVTec-AD, VisA, Real-IAD, MIIC, MVTec-LOCO, and BMAD—under a unified preprocessing and evaluation layer.

The intended scope is broader than standard unsupervised industrial defect detection. MoViAD explicitly supports continual, semi-supervised, few-shot, noisy, Edge, and IoT settings, rather than treating these as afterthoughts. Its stated aim is to be usable by machine learning engineers for custom models, datasets, and backbones, while retaining the flexibility researchers need for new method development (Barusco et al., 16 Jul 2025).

3. Scenario model, method families, and formalization

MoViAD formalizes “unsupervised, noisy” VAD by allowing controlled contamination in the training set. At image level, if $N$ is the number of normal images and $M$ the number of anomalous images, contamination is set through

$\frac{M}{N+M} \approx C.$

At pixel level, with $M_{\text{pxl}}$ anomalous pixels and $N_{\text{pxl}}$ normal pixels, contamination is defined as

$$\frac{M_{\text{pxl}}}{N_{\text{pxl}} + M_{\text{pxl}}}.$$

This lets the same underlying method be stress-tested under realistic training-set impurity rather than only under perfectly clean normal-only assumptions (Barusco et al., 16 Jul 2025).

Continual learning is handled through Replay-based strategies inspired by prior continual VAD work. At task $t$, training includes current-task samples and a replay buffer from previous tasks; the same principle is reported to have been tested with STFPM, VAE, CAE, Ganomaly, FastFlow, and RIAD, with specific replay handling also implemented for PatchCore and PaDiM (Barusco et al., 16 Jul 2025). Supervised, semi-supervised, and few-shot settings are defined primarily through trainer logic and dataset splits rather than by rewriting the core anomaly models.

The method families expose the usual VAD mathematical structure. For PatchCore-style memory banks, a test patch feature $z$ is scored by nearest memory distance,

$s(z) = \min_{m \in \mathcal{M}} \|z - m\|,$

where $\mathcal{M}$ is the bank of normal features. For STFPM-style student–teacher models, training minimizes

$M$0

with anomaly maps derived from teacher–student discrepancies at each spatial location. Reconstruction models minimize losses such as $M$1 or $M$2, and flow-based methods like FastFlow use negative log-likelihood under an invertible feature distribution model (Barusco et al., 16 Jul 2025).

Evaluation is standardized across pixel-level and image-level tasks. Pixel-level metrics include ROC-AUC, PR-AUC, F1-score, and PRO; image-level metrics include ROC-AUC, PR-AUC, and F1-score. Because these are implemented centrally, the same evaluation code can be reused across datasets and model families (Barusco et al., 16 Jul 2025).

4. Backbones, Edge and IoT deployment, and profiling

MoViAD’s backbone layer is designed to make accuracy–efficiency trade-offs explicit. Supported feature extractors include ResNet18, WideResNet, MobileNetV1, MobileNetV2, PhiNet, MCUNet, and MicroNet, with hooks for intermediate multi-scale feature maps and compatibility with PyTorch Hub and Hugging Face Hub (Barusco et al., 16 Jul 2025). This is significant because many VAD methods are backbone-sensitive: memory-bank and student–teacher systems depend directly on the geometry of extracted features.

The Edge scenario assumes the full VAD pipeline runs on a resource-constrained device. MoViAD addresses this through tiny backbones, quantization-aware training, post-training quantization, and compression utilities. The IoT scenario instead splits the pipeline between an edge node and a cloud server: feature extraction can run on-device, features can be quantized or compressed, and the anomaly detector can complete inference server-side while bitrate and communication cost are profiled (Barusco et al., 16 Jul 2025).

The library’s profiling utilities track model size, memory usage, and FLOPs, and its utility layer includes anomaly-map postprocessing, synthetic anomaly generation such as CutPaste, noise injection, spatial deformation, WandB logging, and VLM access via OLLAMA (Barusco et al., 16 Jul 2025). A plausible implication is that MoViAD is intended not only as a benchmark wrapper but as a deployment-oriented experimentation substrate for VAD systems that must move between laboratory, factory, and embedded environments.

5. Cognate anomaly systems in video and streaming

MOVAD, “Memory-augmented Online Video Anomaly Detection,” is an online dashcam anomaly detector for autonomous-driving scenarios. It combines a Short-Term Memory Module implemented by a Video Swin Transformer with a Long-Term Memory Module implemented as a stacked LSTM inside the classifier head, producing a frame-level anomaly score $M$3 without access to future frames. On DoTA, MOVAD reaches an AUC score of $M$4, surpassing the previous state of the art by $M$5 AUC, while using only RGB frames (Rossi et al., 2023).

MMVIAD, “Multi-view Multi-task Video Industrial Anomaly Detection,” defines a different branch of the family: a continuous multi-view video dataset and benchmark for industrial anomaly understanding. It contains object-centric 2-second inspection clips with approximately $M$6 degrees of camera motion, covers $M$7 object categories, $M$8 environments, and $M$9 structural anomaly types, and supports anomaly detection, defect classification, object classification, and anomaly visible-time localization. In the MMVIAD-Unseen protocol, the proposed VISTA model improves the base model’s average score across the four tasks from $\frac{M}{N+M} \approx C.$0 to $\frac{M}{N+M} \approx C.$1, surpassing GPT-5.4 (Zhao et al., 11 May 2026).

MVAD, “Multiple Visual Artifact Detector,” addresses multi-label artifact detection for video streaming rather than semantic anomaly understanding. Its architecture combines an Artifact-aware Dynamic Feature Extractor with a Recurrent Memory Vision Transformer, is trained on a $\frac{M}{N+M} \approx C.$2-patch database generated by streaming-pipeline simulation and Adversarial Data Augmentation, and detects $\frac{M}{N+M} \approx C.$3 target visual artifacts on Maxwell and BVI-Artifact with consistent gains over seven existing detectors (Feng et al., 2024). Together, these systems show that the MoViAD/MOVAD/MMVIAD/MVAD naming cluster spans at least three technical regimes: online traffic anomaly detection, industrial multi-view anomaly reasoning, and artifact-aware streaming-quality diagnosis.

6. MoViAD-style extensions beyond anomaly libraries

In medical imaging, the CMViM paper explicitly frames its method as a prototype of a “Multi-modal Vision Mamba for Alzheimer’s Disease” system. CMViM introduces a 3D Vision Mamba masked autoencoder for paired T1-MRI and PET volumes, augments reconstruction with intra-modal and inter-modal contrastive learning, uses $\frac{M}{N+M} \approx C.$4 inputs split into $\frac{M}{N+M} \approx C.$5 patches, and reports $\frac{M}{N+M} \approx C.$6 accuracy, $\frac{M}{N+M} \approx C.$7 AUC, and $\frac{M}{N+M} \approx C.$8 F1-score on AD classification, yielding at least $\frac{M}{N+M} \approx C.$9 AUC improvement over the strongest baseline (Yang et al., 2024). The paper is explicit that “MoViAD” is not the system’s official name, but it treats the label as a coherent shorthand for a Vision-Mamba-based multi-modal Alzheimer’s disease pipeline.

In human motion analysis, the MoVi dataset paper uses “MoVi-based Action Detection / Analysis” as a practical reading of “MoViAD.” MoVi contains $M_{\text{pxl}}$0 subjects, $M_{\text{pxl}}$1 predefined everyday and sport actions plus $M_{\text{pxl}}$2 self-chosen motion, five capture rounds, approximately $M_{\text{pxl}}$3 hours of motion capture data, $M_{\text{pxl}}$4 hours of video data from four points of view, and $M_{\text{pxl}}$5 hours of IMU data, together with synchronized or calibrated multi-modal annotations including V3D skeletons and MoSh++/AMASS body models (Ghorbani et al., 2020). The term again is not official, but the dataset is presented as directly useful for pose estimation, action recognition, action segmentation, motion prediction, gait analysis, and cross-modal learning.

These adjacent interpretations indicate that “MoViAD” functions as both a concrete library name and a transferable design motif. In its strict sense, it is a modular VAD framework for still-image anomaly detection. In broader research usage, it can denote systems that combine modularity, multi-view or multi-modal perception, and anomaly or action-centric reasoning across industrial, automotive, medical, and human-motion domains (Barusco et al., 16 Jul 2025, Yang et al., 2024, Ghorbani et al., 2020).