Research map / Detection and segmentation / Open-set and few-shot detection
MANTA: A Large-Scale Multi-View and Visual-Text Anomaly Detection Dataset for Tiny Objects
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces MANTA, a multi-view visual-text anomaly detection dataset of 137K five-view images of 38 tiny-object categories with pixel masks, plus declarative-knowledge text and multiple-choice questions, and benchmarks AD methods and a BLIP-2 LoRA baseline.
Problem
Anomaly detection for tiny objects is underexplored: existing datasets are mostly single-view, large industrial objects, and lack text annotations for visual-language anomaly reasoning.
Approach
Capture each tiny object from five fixed viewpoints with a custom prototype rig, pair the visual data with expert-written hierarchical text knowledge (what/why/how) and image-pair multiple-choice questions, and benchmark AD methods under single-view, multi-view, multi-class, text-prompt and visual-language settings.
Key results
- Single-view AD methods reach below 97% I-AUROC on MANTA across domains, with agriculture and groceries averaging below 90% I-AUROC (e.g., PatchCore averages 93.7/95.7 view-eval).
- Multi-view setting averages 91% I-AUROC and 94% P-AUROC across five domains.
- Visual-language baseline (BLIP-2 + LoRA) reaches 52.7% option accuracy and 3.0% question accuracy on ConsL in the zero/few-shot comparison, versus 47-52% for other VLMs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use five-view images instead of single-view images for training | Visual anomaly detection methods on MANTA · MANTA visual, average across domains | I-AUROC | Helped |
| Train one model across multiple categories rather than one-for-one | Visual anomaly detection methods on MANTA · MANTA visual, average across domains | I-AUROC | Hurt |
| Use DeclK text prompts with WinCLIP for one-shot anomaly detection | WinCLIP · MANTA DeclK, one-shot, average across domains | I-AUROC: 76.7 | Helped |
| Use few-shot training rather than zero-shot inference | Visual-language models on MANTA ConsL · MANTA ConsL | option accuracy: 55% | Helped |
| Use few-shot training rather than zero-shot inference | Visual-language models on MANTA ConsL · MANTA ConsL | question accuracy: 2% | Helped |
| Use multi-view inputs | Visual anomaly detection methods on MANTA · MANTA visual, single-view versus multi-view | localization consistency | Mixed |
Surprising findings
- Reconstruction-based methods performed worse than other approaches, attributed to large anomalous regions in tiny objects.
- Multi-class models fell below 90% I-AUROC, lower than single-view and multi-view settings, due to variability among objects within a domain.
Limitations
- Five cameras leave some surfaces (e.g., hexagonal sides) partially uncovered.
- Text data subset is small due to cost constraints.
- Visual-language baseline performs poorly, attributed to domain gap with pretrained ViT data.
Experimental setup
Datasets: MANTA
Benchmarks: MANTA
Metrics: I-AUROC P-AUROC option accuracy question accuracy
Baselines: RD PatchCore CDO DMAD SimpleNet UniAD CRAD HGAD WinCLIP PromptAD VCP-CLIP LLaVA-1.5
Builds on: BLIP-2 PatchCore WinCLIP RD
Closest papers by meaning
- Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.86)
- Unseen Visual Anomaly Generation (CVPR 2025 · similarity 0.86)
- AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language Models (AAAI 2024 · similarity 0.86)
- LogicAD: Explainable Anomaly Detection via VLM-based Text Feature Extraction (AAAI 2025 · similarity 0.85)
- UniVAD: A Training-free Unified Model for Few-shot Visual Anomaly Detection (CVPR 2025 · similarity 0.85)
- One-to-Normal: Anomaly Personalization for Few-shot Anomaly Detection (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.