ImageNet 数据集发布

在 ImageNet 能帮系统认出一张图之前，得先有人决定这张图叫什么。

互联网上已经有数以十亿计的照片。搜索引擎能捞出候选，一堆文件却不会自己长成训练集。「哈士奇」「帆船」「乐器」要成为类别，先得有人规定这些名字之间是什么关系，再逐张判断搜索结果究竟算不算数。

李飞飞团队借用了 WordNet。这个英语词汇数据库把名词组织成一层层同义词集（synset）；ImageNet 试着给这些节点填入图像。选择 WordNet 而不是另起一套任意标签列表，是为了让视觉类别与语言中的名词层级对齐，方便扩展到数以万计的概念。分类树先于模型存在：哺乳动物下面有食肉目，犬科下面还有更细的犬种。2009 年 CVPR 论文 ImageNet: A Large-Scale Hierarchical Image Database 报告的版本覆盖十二棵子树，共 5,247 个 synset、约 320 万张经过清理的图片，平均每类六百多张。论文设想的终点要大得多——为 WordNet 中数以万计的名词概念各找数百到上千张图；后来公开叙述中的「超过 1400 万张」是持续扩充后的规模，不应与 2009 年论文中的数字混为一次发布。

搜索引擎负责把图捞上来，可靠性则要靠人。候选图片被送到 Amazon Mechanical Turk，由众包标注者核验一张图是否属于提议的 synset。标注界面通常给出类别名与示例，工人做的是核验与筛选，而不是自由发明分类体系——体系早已由 WordNet 与项目团队定好。这个环节很少出现在后来那些漂亮的准确率曲线里，却决定了曲线有没有共同的横轴。所谓「大数据」在这里并不是自动积累的自然资源，而是一连串选择：用什么查询词，哪些图片留下，一张图片归在哪个节点，多少人的判断才足以确认标签。

ImageNet 此后继续扩充。2010 年启动的年度大规模视觉识别竞赛（ILSVRC）从中划出约 120 万张训练图像和 1,000 个类别。研究者第一次可以在同一批图片、同一套指标（如 top-1 / top-5 错误率）和同一个截止日期前交卷。算法之间的差异不再被各自挑选的数据集遮住。2012 年 AlexNet 那个异常醒目的领先，也因此能被清楚地量出来：差距写在共享的测量仪器上。

统一尺度的代价同样藏在分类表里。WordNet 的名词体系是英语词汇组织，不是世界的中性副本；网络图片也不是均匀抽取的现实。类别边界、地域分布、人物相关标签、有争议的社会类别，以及来源不明的版权史，都会跟着数据进入模型。后来关于数据集偏见、人脸与人物标签、以及网络爬取图像的权利问题，都部分回溯到这类「公共研究基础设施」如何被建造。

在像素进入卷积层之前，目录已经决定了世界的切法。后来的模型可以换架构、换损失函数、换预训练配方；若仍在 ImageNet 或其衍生协议上报告分数，它们仍在同一张目录表上争名次。目录的作者权，比某一届冠军网络更长久。

Before ImageNet could help a system recognize an image, someone had to decide what the image would be called.

The web already held billions of photographs. Search engines could retrieve candidates, but a pile of files was not a training set. A category such as “husky,” “sailboat,” or “musical instrument” needed a place among other categories, and every result needed a judgment about whether it belonged.

Fei-Fei Li’s team borrowed that order from WordNet, the lexical database that arranges English nouns into hierarchical synonym sets. ImageNet aimed to populate these nodes with pictures. Choosing WordNet rather than an arbitrary flat label list aligned visual categories with a linguistic hierarchy, making it easier to scale toward tens of thousands of concepts. Mammals divided into narrower branches; dogs divided again into breeds. The 2009 CVPR paper ImageNet: A Large-Scale Hierarchical Image Database described twelve subtrees, 5,247 synsets, and about 3.2 million cleaned images, averaging more than six hundred per category. Its stated ambition reached much further: hundreds or thousands of examples for tens of thousands of WordNet noun concepts. The later public figure of more than fourteen million images reflects continued growth; it should not be collapsed into the 2009 paper’s release snapshot.

Search engines brought candidate images into view. Human labelers made the collection trustworthy enough to use. Tasks were distributed through Amazon Mechanical Turk, where workers verified whether an image matched a proposed synset. The interface typically offered a category name and examples; workers verified and filtered rather than inventing the taxonomy freely—WordNet and the project team had already fixed the system of names. That labor rarely appears in the smooth accuracy curves published later. Yet it established the common x-axis beneath them. “Large-scale data” was not a natural deposit awaiting extraction. It was a succession of choices about queries, inclusion, category, and sufficient agreement.

ImageNet continued to grow beyond the version described in 2009. The annual ImageNet Large Scale Visual Recognition Challenge, beginning in 2010, selected a competition set of roughly 1.2 million training images across 1,000 categories. Researchers now met the same data, metrics (including top-1 and top-5 error), and deadline. Differences between algorithms could no longer hide as easily behind different test collections. When AlexNet opened an unusually large performance gap in 2012, the gap was legible because the measuring instrument had already been built on a shared public subset.

The instrument contained its own assumptions. WordNet is an English lexical organization, not a neutral copy of the world. Web images are not a uniform sample of human experience. Category boundaries, geographic distributions, labels applied to people, contested social categories, and uncertain copyright histories all travel from the dataset into the models trained on it. Later debates about dataset bias, face and person labels, and the rights status of crawled images partly retrace how this “public research infrastructure” was built.

The catalog cuts the world before the convolution begins. Later models may change architectures, losses, and pretraining recipes; if they still report scores under ImageNet or its derived protocols, they are still contesting places on the same table of names. Authorship of the catalog outlasts any one champion network.