OPC16K
收藏资源简介:
OPC16K是由南京大学和伦敦帝国理工学院联合构建的大规模真实伪装目标检测基准数据集,旨在突破传统封闭世界假设,系统评估模型在开放场景下的性能。该数据集包含16,245张图像,来自14个来源,精心划分为9,000张伪装目标图像、3,050张纯背景图像和4,195张非伪装目标图像,兼顾了分割质量与负样本抑制能力的评估。在构建过程中,通过基于伪装目标类别分布的两轮收集策略,实现了正负样本在场景与类别上的分布对齐,避免模型利用领域差异作为捷径。该数据集主要应用于伪装目标检测领域,解决现有模型在真实部署场景中对纯背景或非伪装目标图像产生大量误报的问题,推动模型从纯分割任务向目标存在性推理的转变。
OPC16K is a large-scale real-world camouflaged object detection benchmark dataset jointly developed by Nanjing University and Imperial College London, aiming to break through the traditional closed-world assumption and systematically evaluate the performance of models in open scenarios. This dataset contains 16,245 images from 14 sources, which are carefully divided into 9,000 camouflaged object images, 3,050 pure background images, and 4,195 non-camouflaged object images, enabling simultaneous evaluation of both segmentation quality and negative sample suppression capability. During the construction process, a two-round collection strategy based on the category distribution of camouflaged objects was adopted to align the distribution of positive and negative samples in terms of scenarios and categories, preventing models from leveraging domain differences as shortcuts. This dataset is primarily applied in the field of camouflaged object detection, addressing the issue that existing models generate a large number of false alarms on pure background or non-camouflaged object images in real-world deployment scenarios, and promoting the transition of models from pure segmentation tasks to object existence reasoning.
数据集概述
该数据集名为 OPCOD,全称为 Over-optimistic Camouflaged Object Detection Dataset,由论文 Is There Really a Camouflaged Object? Towards Realistic Camouflaged Object Detection 提出。
核心目标
该数据集针对现实场景中的伪装目标检测任务构建,旨在解决现有伪装目标检测数据集中存在的“过度乐观”问题,使模型能够处理更真实、更具挑战性的伪装场景。
关键特性
- 数据构成:包含真实场景中的伪装目标图像,其伪装程度和背景复杂度高于传统基准数据集。
- 任务类型:主要用于伪装目标检测(Camouflaged Object Detection, COD)的模型训练与性能评估。
- 研究意义:通过引入更贴近实际应用的样本,推动伪装目标检测从理想化设定向现实应用场景过渡。
补充说明
- 数据集的代码预计将在近期开源(README中注明“We will open-source the code as soon as possible”)。
- 数据集详情页地址:
https://github.com/2231122/OPCOD

- 1Is There Really a Camouflaged Object? Towards Realistic Camouflaged Object Detection南京大学; 伦敦帝国理工学院 · 2026年



