conamodal
收藏资源简介:
ConAmodal是一个专为建筑工地场景设计的非模态实例分割(amodal instance segmentation)基准数据集。其目标是在建筑施工环境中,恢复被其他物体部分遮挡的结构元素(如梁、柱、钢筋、管道)的完整轮廓。与现有的非模态分割数据集(如COCOA、KINS、D2SA)主要覆盖日常场景(人物、车辆、家居物品)不同,ConAmodal聚焦于建筑元素的独特属性——它们通常呈长条状、直线形、重复性强且具有强烈的形状先验。该数据集包含1643个实例,共4925个样本,涵盖三个主要类别:structural_member(结构构件)、piping(管道)和rebar(钢筋),并进一步细分为12个子类别。数据集由三个子集构成:conamodal_generation2(用于训练和模型选择,包含合成遮挡和AI生成图像)、conamodal_field(作为测试基准,包含合成遮挡和来自5个真实建筑工地的照片)、以及conamodal_field_realocc(包含真实遮挡,但仅用于稳定性检查,不作为基准的一部分)。每个样本提供遮挡后的图像、真实非模态掩码(完整形状)、真实模态掩码(可见部分),以及遮挡前的图像(如果存在)作为生成方法的RGB目标。设计上,ConAmodal通过仅接受自然遮挡程度≤5%的对象来保证非模态真值的准确性,然后以可控速率和深度感知方式添加合成遮挡,从而使得遮挡率、遮挡物数量和碎片数量可作为难度评估指标。由于数据规模相对较小,子类别级别的统计结果仅供参考;合成遮挡与真实遮挡存在差异;真实遮挡的真值基于重建而非直接观测;训练图像为AI生成,与真实照片存在领域差距。数据集采用CC-BY-NC-4.0非商业许可证。
ConAmodal is an amodal instance segmentation benchmark dataset specifically designed for construction site scenarios. Its goal is to recover the complete contours of structural elements (e.g., beams, columns, rebar, pipes) that are partially occluded by other objects in building construction environments. Unlike existing amodal segmentation datasets (such as COCOA, KINS, D2SA) which mainly cover everyday scenes (people, vehicles, household items), ConAmodal focuses on the unique attributes of building elements—they are typically long, linear, repetitive, and have strong shape priors. The dataset contains 1,643 instances and 4,925 samples, covering three main categories: structural_member, piping, and rebar, further subdivided into 12 subcategories. The dataset consists of three subsets: conamodal_generation2 (for training and model selection, containing synthetic occlusions and AI-generated images), conamodal_field (as a test benchmark, containing synthetic occlusions and photos from 5 real construction sites), and conamodal_field_realocc (containing real occlusions, but only for stability checks, not part of the benchmark). Each sample provides an occluded image, a ground-truth amodal mask (full shape), a ground-truth modal mask (visible part), and an image before occlusion (if available) as the RGB target for generation methods. By design, ConAmodal ensures the accuracy of amodal ground truth by only accepting objects with natural occlusion ≤5%, then adding synthetic occlusions at controlled rates and depth-aware manners, so that occlusion rate, number of occluders, and fragment count can serve as difficulty evaluation metrics. Due to the relatively small data scale, subcategory-level statistical results are for reference only; synthetic occlusions differ from real occlusions; ground truth for real occlusions is based on reconstruction rather than direct observation; training images are AI-generated and have a domain gap with real photos. The dataset is licensed under CC-BY-NC-4.0 (non-commercial).
ConAmodal 数据集
概述
ConAmodal 是一个用于建筑施工场地模态实例分割的基准数据集,目标是恢复被其他物体部分遮挡的结构元素(如梁、柱、钢筋、管道)的完整轮廓。与现有的日常场景模态基准(COCOA、KINS、D2SA)不同,本数据集专注于建筑元素——这些元素具有长、直、重复且遵循强形状先验的特点。
数据集规模与构成
- 规模:约 1K-10K 样本,包含 4,925 个样本、1,643 个实例
- 主要基准:包含
conamodal_generation2(合成图像,用于训练和模型选择)和conamodal_field(真实图像,5 个施工现场,作为主测试集) - 附加集:
conamodal_field_realocc为真实遮挡数据,不属于 ConAmodal 基准,用于野外一致性检查
类别
- 3 个主要类别:
structural_member(结构构件)、piping(管道)、rebar(钢筋) - 包含 12 个子类别
数据特点
- 每个样本提供:遮挡图像、真值模态掩码(完整形状)、真值可见掩码(可见部分)、以及遮挡前图像(作为生成方法的 RGB 目标)
- 真值获取策略:标注者仅接受自然遮挡 ≤5% 的几乎完全可见对象,避免猜测性标注;遮挡以受控速率合成添加,遮挡物按类别自然尺寸先验缩放并深度感知放置
- 难度轴:遮挡率、遮挡物数量、碎片数量可作为难度评估维度
使用注意事项
- Windows 文件系统无法直接下载(文件名含
::) - 合成集中的每个实例在轻/中/重遮挡下出现三次,共享同一
ann_id,划分数据时必须使用提供的splits/generation2/split_manifest.json,避免样本级划分导致同一物理对象泄漏到两侧 conamodal_field和conamodal_field_realocc是留出评估集,应整体使用,不可用于调参
已知局限
- 规模较小,子类别结果的统计功效较低
- 主要集采用合成遮挡,与真实遮挡存在差异
- 真实遮挡真值为重建结果而非观测结果,仅适用于排名稳定性检查
- 每个施工场地仅在一天拍摄,存在场地和进度偏差,一个场地无管道图像
- 生成的训练图像与真实照片存在域差距
文档
- SCHEMA.md:布局、元数据字段、划分规则
- DATASHEET.md:动机、收集、标注、用途、局限性
- LICENSE:组件级许可
- LICENSES-THIRD-PARTY.md:第三方材料和归属要求
- croissant.json:机器可读元数据(ML Commons Croissant 1.0)
- SHA256SUMS:完整性校验和
许可
组件特定许可,组合数据集为非商业用途(CC-BY-NC-4.0)。





