---
title: CoalAD
canonical_url: "https://www.modelscope.cn/datasets/lyfjwp/CoalAD"
md_url: "https://www.modelscope.cn/datasets/lyfjwp/CoalAD.md"
repository: lyfjwp/CoalAD
last_updated: 2026-03-12
license: CC-BY-4.0
storage_size: "5.8 GB"
downloads: 1356
stars: 0
---

# CoalAD

> CoalAD - lyfjwp 在 ModelScope 开源的数据集。CoalAD

lyfjwp/CoalAD 是 ModelScope 魔搭社区上的数据集，存储大小 5.8 GB，采用 CC-BY-4.0 许可。

- **Repository**: lyfjwp/CoalAD
- **License**: CC-BY-4.0
- **Storage size**: 5.8 GB
- **Downloads**: 1356
- **Stars**: 0
- **Last updated**: 2026-03-12

Source: https://www.modelscope.cn/datasets/lyfjwp/CoalAD

---

## Dataset Description (CoalAD)
CoalAD is a benchmark for **unsupervised foreign-object anomaly detection and pixel-level localization** in conveyor-belt coal-stream scenes. It is curated from the public DsCGF dataset collected in a real coal preparation plant. We select the overhead conveyor-view subset containing foreign objects and apply preprocessing (duplicate removal, normal/anomalous split, and unified resizing) to build CoalAD.

**Task setting:** models are trained using **normal-only** images (coal and gangue only) and are required to detect and localize unknown foreign objects at test time (e.g., wood, metal parts, nets/ropes, plastic bags). This scenario is highly unstructured: coal/gangue are randomly piled, the background is complex and time-varying, and foreign objects often become low-contrast and boundary-blurred due to compression, staining, dust contamination, and occlusion—making many methods designed for structured industrial settings degrade noticeably.

**Annotations:** binary anomaly masks are generated from DsCGF instance segmentation annotations (foreign objects = 1, background = 0) and manually reviewed for consistency.

**Splits:** 2,490 normal images for training; the test set contains 811 normal and 943 anomalous images. Wooden objects dominate the anomalous set (799 images), while other anomalies include nets, metal braces/rods, gloves, and plastic bags.

### Examples
- **example1.png:** normal vs. low-contrast anomalies (a–d: normal; e–h: anomalous), illustrating unstructured characteristics and blurred boundaries under low contrast/occlusion/discoloration.  
  ![CoalAD examples](https://www.modelscope.cn/datasets/lyfjwp/CoalAD/resolve/master/example1.png)

- **example2.png:** representative foreign objects (left: wood; right: others such as nets and metal parts). Even within wood, appearances vary substantially due to compression, staining, and surface contamination.  
  ![CoalAD objects](https://www.modelscope.cn/datasets/lyfjwp/CoalAD/resolve/master/example2.png)

## Citation
If you find this dataset helpful, please consider citing our paper:

如果该数据集对您的研究有帮助，欢迎引用我们的论文：
```bash
@misc{CoalAD,
      title={Semantic-Deviation-Anchored Multi-Branch Fusion for Unsupervised Anomaly Detection and Localization in Unstructured Conveyor-Belt Coal Scenes}, 
}
```
For the dataset file metadata and the data files, please visit the **“数据集文件”** (Dataset Files) page.

数据集文件元信息以及数据文件，请浏览“数据集文件”页面获取。  

## 数据集简介（CoalAD）
CoalAD 是一个面向**输送带煤流场景**的无监督异物异常检测与像素级定位基准。数据源自 Lv 等人发布的 DsCGF（真实选煤厂采集），我们筛选了俯视输送带视角下包含异物的子集，并进行了去重、正常/异常划分与统一分辨率等预处理，形成 CoalAD。

**任务设定：**训练阶段仅使用正常样本（仅含煤与矸石）；测试阶段需要检测并像素级定位未知异物（如木料、金属件、兜网/绳索、塑料袋等）。该场景高度非结构化：煤/矸石随机堆叠、背景复杂且随时间变化，异物常因挤压、染色、煤尘污染与遮挡而呈现低对比度与模糊边界，使得结构化工业场景中常见方法容易退化。

**标注：**基于 DsCGF 的实例分割标注生成二值 mask（异物为 1，其余为 0），并进行人工复核以提升一致性。

**数据划分：**训练集 2,490 张正常图像；测试集包含 811 张正常与 943 张异常。异常中木料占比最高（799 张），其余包含兜网、金属支架/杆件、手套、塑料袋等。

### 示例
- **example1.png：**正常样本与低对比度异常示例（左 a–d 为正常，右 e–h 为异常），展示非结构化堆叠与低对比度、遮挡、染色导致的模糊边界。  
  ![CoalAD examples](https://www.modelscope.cn/datasets/lyfjwp/CoalAD/resolve/master/example1.png)

- **example2.png：**异物类型示例（左侧为木料，右侧为其他类型如兜网与金属件）；即使同为木料，也会因挤压、染色与表面污染而外观差异显著。  
  ![CoalAD objects](https://www.modelscope.cn/datasets/lyfjwp/CoalAD/resolve/master/example2.png)

---

#### 下载方法 
:modelscope-code[]{type="sdk"}
:modelscope-code[]{type="git"}