| --- |
| datasets: |
| - geonuk-kimmm/Inspect-Anything |
| pretty_name: Inspect Anything |
| license: mit |
| task_categories: |
| - object-detection |
| tags: |
| - Open-set |
| - Visual-Prompt |
| - Visual-Inspection |
| - Smart-Factory |
| - Defect-Segmentation |
| language: |
| - en |
| --- |
| |
| # Inspect Anything |
|
|
| This repository provides the Inspect-Anything (InsA) benchmark **annotation files (JSON)** proposed in our paper, without the original images. |
| The InsA benchmark is introduced in the **UniSpector: Towards Universal Open-set Defect Recognition via Spectral-Contrastive Visual Prompting (CVPR 2026)**. |
| For instructions on how to prepare the images, see the [dataset preparation](https://github.com/geonuk-kimmm/UniSpector/?tab=readme-ov-file#dataset-preparation) section. |
|
|
|
|
|
|
| --- |
|
|
| ## Overview |
| To evaluate generalization ability to unseen defects, we introduce **Inspect Anything (InsA)**, a benchmark for open-set defect detection and segmentation within a visual prompting framework. |
|
|
| InsA evaluates three aspects of generalization: |
| - (1) in-domain generalization to novel defect types that share visual similarity with the seen sets, |
| - (2) cross-domain transferability to unseen defects emerging under different material properties, imaging conditions, and defect morphologies, and |
| - (3) prompt-level generalization, quantified by aggregating results across multiple class partitions and independently sampled prompt sets to reflect stability rather than dependence on a single prompt configuration. |
|
|
| We construct this benchmark from seven industrial inspection datasets: `GC10-DET`, `Magnetic Tile Surface Defect`, `Real-IAD`, `MVTec AD`, `3CAD`, `VISION`, and `VisA`. |
| Among these, `GC10-DET`, `Magnetic Tile Surface Defect`, `Real-IAD`, and `MVTec AD` are used as **in‑domain** datasets, with defect categories partitioned into **seen** classes (training) and **unseen** classes (testing). |
| In contrast, `3CAD`, `VISION`, and `VisA` serve as **cross‑domain** datasets, used exclusively for evaluating transferability to novel defect types and object appearances. |
| To ensure robust estimates of generalization performance, we form **three independent seen–unseen splits** using random seeds `{42, 82, 777}`, holding out roughly **25% of defect categories** in each in‑domain dataset as **unseen** test classes. |
|
|
| --- |
|
|
| ## Dataset Composition and Characteristics |
|
|
| The InsA benchmark is composed of diverse products and materials, enabling evaluation of **open‑set visual inspection** performance across varied distributions. |
| Defect images are captured under a mix of distinct imaging conditions (e.g., very bright or very dark illumination), resulting in generally large standard deviations in HSV channels. |
| These distributions also vary across datasets, leading to differing mean values and further highlighting the challenge of generalizing visual inspection models across heterogeneous industrial data. |
|
|
|
|
| | Dataset | Images | Defect Instances | Defect Categories | In-domain Seen (Train) | In-domain Unseen (Test) | Cross-domain | |
| | :---: | :---: | :---: | :---: | :---: | :---: | :---: | |
| | GC10 | 2,292 | 3,563 | 10 | 7 | 3 | - | |
| | MagneticTile | 388 | 514 | 5 | 4 | 1 | - | |
| | Real-IAD | 49,237 | 55,326 | 111 | 83 | 28 | - | |
| | MVTec | 1,258 | 1,836 | 73 | 53 | 20 | - | |
| | 3CAD | 11,074 | 16,559 | 46 | - | - | 46 | |
| | VISION | 1,757 | 3,553 | 44 | - | - | 44 | |
| | VisA | 1,167 | 2,131 | 71 | - | - | 71 | |
| | Total | 67,173 | 83,482 | 360 | 147 | 52 | 161 | |
|
|
|
|
| The table below summarizes key characteristics of each dataset, including product types, materials, and color distribution statistics (mean ± std) for each HSV channel. |
|
|
| | Dataset | Product | Material | H (mean ± std) | S (mean ± std) | V (mean ± std) | |
| | :---: | :---: | :---: | :---: | :---: | :---: | |
| | GC10‑DET | Steel | Steel | `0.0 ± 0.0` | `0.0 ± 0.0` | `85.7 ± 39.6` | |
| | Magnetic Tile | Magnetic Tile | Steel | `0.0 ± 0.0` | `0.0 ± 0.0` | `109.4 ± 47.4` | |
| | Real‑IAD | PCB, Toy Brick, Transistor, etc. | Plastic, Rubber, Wood, etc. | `33.4 ± 49.5` | `40.6 ± 65.7` | `110.2 ± 105.5` | |
| | MVTec AD | Cable, Hazelnut, Tile, etc. | Glass, Metal, Fabric, etc. | `46.7 ± 55.1` | `46.0 ± 47.8` | `120.5 ± 66.7` | |
| | 3CAD | Camera Cover, Tablet PC, etc. | Aluminum, Copper, etc. | `7.2 ± 19.2` | `16.7 ± 49.0` | `78.6 ± 70.5` | |
| | VISION | Capacitor, Lens, Screw, etc. | Plastic, Steel, Wood, etc. | `25.2 ± 41.1` | `38.4 ± 70.7` | `106.3 ± 83.3` | |
| | VisA | Candle, Capsule, Macaroni, etc. | Plastic, Food, etc. | `47.7 ± 36.4` | `95.7 ± 71.9` | `101.9 ± 66.1` | |
|
|
| --- |
|
|
| ## Detailed Description |
|
|
|
|
| ### GC10‑DET |
|
|
| `GC10-DET` is a defect detection dataset collected from metallic surfaces in industrial environments. It contains about **2,300 defect images** with bounding‑box annotations for **10 defect categories**: |
| `silk spot`, `welding line`, `punching hole`, `water spot`, `crescent gap`, `oil spot`, `inclusion`, `waist folding`, `crease`, and `rolled pit`. |
|
|
|
|
|
|
| ### Magnetic Tile Surface Defect |
|
|
| The magnetic tile surface defect dataset contains images of magnetic tile surfaces collected from industrial production lines, covering both defective and defect‑free samples. |
| It provides **pixel‑level annotations** for **5 defect categories**: `blowhole`, `crack`, `break`, `fray`, and `uneven`. |
| In our benchmark, we discard normal images and retain only defective samples, keeping approximately **400 defect images** from the original release. |
|
|
|
|
|
|
| ### Real‑IAD |
|
|
| `Real-IAD` is a multi‑view industrial anomaly dataset comprising **30 real‑world objects** fabricated from diverse materials (e.g., plastic, rubber). |
| Each object is captured under **5 viewpoints** and exhibits **2–5 distinct defect modes** selected from **8 defect families**: |
| `pit`, `deformation`, `abrasion`, `scratch`, `damage`, `missing parts`, `foreign objects`, and `contamination`. We keep only viewpoints equipped with **pixel‑level defect masks** as anomaly samples and use the official **1024×1024** resolution images. |
| From the binary defect masks, we perform **connected‑component labeling** with 8‑connectivity to obtain instance‑level segments, and discard components whose width or height is smaller than **1%** of the image width or height to remove tiny noisy polygons. |
|
|
|
|
|
|
| ### MVTec AD |
|
|
| `MVTec AD` is a real‑world industrial anomaly dataset that contains defective images across multiple object and texture categories, with various defect types such as contamination, oil, cuts, and cracks. |
| In our benchmark, we discard all “good” images and use only **anomalous samples with polygon masks**. From these masks, we extract individual defect instances and remove components whose width or height is smaller than **1%** of the corresponding image dimension to filter out tiny noisy regions. |
|
|
|
|
| ### 3CAD |
|
|
| `3CAD` is a large‑scale anomaly detection dataset collected from real 3C product manufacturing lines, covering representative defects that arise in practical production environments. |
| It focuses on parts made of three common materials (**Aluminum, Iron, Copper**) and includes multiple types of 3C components (e.g., camera covers, tablets, and PCs). Since 3CAD is designed for anomaly detection, we discard the “good” class (without polygon annotations) and also remove the **`Multiple-defects`** class, whose defect instances cannot be clearly assigned to specific categories. |
| After this filtering, we obtain a total of **46 defect categories**. From the remaining polygons, we extract individual defect instances and discard components whose width or height is smaller than **1%** of the image dimension to remove tiny noisy regions. |
|
|
|
|
|
|
| ### VISION |
|
|
| `VISION` is a benchmark unifying **14 industrial inspection subsets**, each corresponding to a distinct object class from real manufacturing lines and captured at its native (often high) resolution; image sizes therefore vary across subsets. |
| It provides **pixel‑level instance masks** for **44 defect categories**. For InsA, we retain only images in the `train` and `val` partitions that include polygon annotations and discard the `inference` split, whose labels are withheld. |
| From the polygon masks, we remove segments whose width or height is smaller than **1%** of the image width or height to filter out tiny noisy regions. |
|
|
|
|
|
|
| ### VisA |
|
|
| `VisA` is an industrial visual inspection dataset containing normal and defective images from **12 object categories**, some of which exhibit large variations in object location and pose across images. |
| It covers both **surface‑level defects** (e.g., scratches, dents) and **structural defects** (e.g., misplacement). In our benchmark, we discard all **normal images** and retain only those that contain at least one annotated defective region. |
| We also exclude the **`Other`** defect class, whose semantics are not clearly specified. From the binary defect masks, we apply **connected‑component labeling** with 8‑connectivity to obtain instance‑level segments, and discard components whose width or height is smaller than **1%** of the image dimension to remove tiny noisy regions. |
|
|
| --- |
|
|
| ## Acknowledgement |
|
|
| We gratefully acknowledge [Minhoi Kim](https://www.linkedin.com/in/%EB%AF%BC%ED%9A%8C-%EA%B9%80-081435253/) for their major contributions to the construction and refinement of this dataset. |
|
|
| --- |
|
|
| ## License |
|
|
| This dataset is released under the MIT License, which allows unrestricted use for both non-commercial and commercial purposes. |
