ConceptSeg-Benchmark
收藏资源简介:
该数据集是一个层次化概念分割基准套件,用于评估ConceptSeg-R1模型在三个认知层次(上下文无关概念、上下文相关概念和上下文推理概念)上的分布外泛化能力。它通过统一分类法整合了多个现有公开数据集,支持规则归纳训练和结构化评估。数据内容涵盖广泛的视觉概念和场景,包括来自COCO数据集的常见生活对象(如人、动物)和人造物品(如车辆、工具),来自iNaturalist的长尾和稀有类别,以及来自多个专业领域的数据:如显著性检测(DUTS)、伪装物体检测(COD10K)、透明物体分割(Trans10K)、阴影检测(SBU)、工业表面缺陷检测(ESDIs-SOD)、医学图像分析(如结肠息肉、乳腺超声肿瘤、皮肤病变)以及需要复杂推理的多图像 grounding 任务(MGrounding, MIG-Bench)。整个基准套件由15个子数据集构成,为概念分割模型的鲁棒性和泛化能力提供了多层次、多领域的综合评估框架。
This dataset is a hierarchical concept segmentation benchmark suite designed to evaluate the out-of-distribution generalization capabilities of the ConceptSeg-R1 model across three cognitive levels: context-independent (CI) concepts, context-dependent (CD) concepts, and context-reasoning (CR) concepts. It integrates multiple existing public datasets through a unified taxonomy, supporting rule-induction training and structured evaluation. The data covers a wide range of visual concepts and scenes, including common everyday objects (e.g., people, animals) and man-made items (e.g., vehicles, tools) from COCO, long-tail and rare categories from iNaturalist, and data from various specialized domains: such as saliency detection (DUTS), camouflaged object detection (COD10K), transparent object segmentation (Trans10K), shadow detection (SBU), industrial surface defect detection (ESDIs-SOD), medical image analysis (e.g., colon polyps, breast ultrasound tumors, skin lesions), and multi-image grounding tasks requiring complex reasoning (MGrounding, MIG-Bench). The entire benchmark suite consists of 15 sub-datasets, providing a multi-level, multi-domain comprehensive evaluation framework for the robustness and generalization of concept segmentation models.
数据集概述:ConceptSeg-Benchmark(层次化概念分割基准套件)
数据集简介
ConceptSeg-Benchmark 是一个层次化概念分割基准套件,用于评估 ConceptSeg-R1 模型在分布外泛化能力上的表现。该基准涵盖三个认知层次的概念:上下文无关(CI)、上下文依赖(CD) 和 上下文推理(CR)。
基准结构
该套件将多个现有数据集整合到统一的分类体系中,支持规则归纳训练和结构化评估。完整基准组成如下表所示:
| 概念层次 | 目录名称 | 来源数据集 | 对应论文/数据集名称 | 描述 |
|---|---|---|---|---|
| 上下文无关(CI) | coco2014_Living | COCO | One-shot learning for semantic segmentation | 常见生活物体(如人、动物) |
| 上下文无关(CI) | coco2014_Artifact | COCO | One-shot learning for semantic segmentation | 人造物体类别(如车辆、工具) |
| 上下文无关(CI) | ultra_rare | iNaturalist | Sam 3: Segment anything with concepts | 长尾类别(模型置信度底部1%) |
| 上下文推理(CI) | rare | iNaturalist | Sam 3: Segment anything with concepts | 分布外/稀有类别评估 |
| 上下文推理(CI) | fewshot1000 | FSS-1000 | FSS-1000: A 1000-Class Dataset for Few-shot Segmentation | 一致性:识别支持集中的共享模式 |
| 上下文推理(CI) | CoSOD3k1024 | COSOD3K | Co-Salient Object Detection: A Benchmark and Algorithms | 共显著性:推理发现共存物体 |
| 上下文依赖(CD) | DUTS | DUTS | DUTS: A Large-scale Dataset for Salient Object Detection | 显著性:从背景中突出的目标 |
| 上下文依赖(CD) | COD10K1024 | COD10K | COD10K: A Large-scale Camouflaged Object Detection Dataset | 伪装:融入背景的目标 |
| 上下文依赖(CD) | transparent1024 | Trans10K | Trans10K: A Large-scale Dataset for Transparent Object Segmentation | 透明度:材料导致的光学折射 |
| 上下文依赖(CD) | Shadow_detection | SBU | Large-scale training of shadow detectors with noisily-annotated shadow examples | 阴影:上下文交互导致的改变 |
| 上下文依赖(CD) | ESDIDefects | ESDIs-SOD | Autocorrelation aware aggregation network for salient object detection of strip steel surface defects | 工业异常:钢材表面制造缺陷 |
| 上下文依赖(CD) | Polyp | Kvasir/CVC | Pranet: Parallel reverse attention network for polyp segmentation | 医学:结肠息肉病变识别 |
| 上下文依赖(CD) | Breast_Tumor | Dataset-B | Dataset of breast ultrasound images | 医学:乳腺超声肿瘤识别 |
| 上下文依赖(CD) | isic2018 | ISIC2018 | Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic) | 医学:皮肤病变分析(相对于健康组织) |
| 上下文推理(CR) | MGrounding-630k | MGrounding | Multi-Image Grounding for Visual Reasoning | 逻辑:推理识别不同或共享物体 |
| 上下文推理(CR) | MIG-Bench | MIG | Migician: Revealing the magic of free-form multi-image grounding in multimodal large language models | 复杂CR:包含时空性和视角差异 |
三个认知层次说明
- 上下文无关(CI):基于物体固有属性进行分割,不依赖上下文信息。
- 上下文依赖(CD):分割结果依赖于物体与周围环境的交互关系(如显著性、伪装、医学异常等)。
- 上下文推理(CR):需要跨图像或多模态逻辑推理才能完成分割任务(如多图像对位、时空推理)。




