marcelohaps/ijb-a

IJB-A HF-ready 是 IARPA Janus Benchmark-A (IJB-A) 人脸识别数据集的 HuggingFace 版本，采用 CleanData 布局，并包含官方的 10 次分割的 1:1 验证和 1:N 识别协议。与 LFW 风格基准不同，IJB-A 是基于模板的：每个主体由一个聚合了多张静态图像和/或视频帧的模板表示。协议 CSV 文件将每行 (模板, 文件) 映射到包含边界框、关键点（如右眼、左眼、鼻基坐标）和人口统计属性（如性别、肤色、年龄、面部毛发）的人脸标注。数据集包含 5396 张静态图像和 20369 个视频帧，总计 25765 张图像，但协议引用了 25791 个独立文件（其中 26 个视频帧缺失）。数据布局包括图像文件夹、打包的视频帧归档文件、协议 CSV 文件（用于训练、验证和识别）以及基线输出文件。此外，还提供了 files.csv 用于快速索引和检查文件覆盖情况。使用前需解压 frame.tar 归档文件以访问视频帧。数据集适用于图像分类和零样本图像分类任务，特别关注人脸识别、验证和识别场景。IJB-A HF-ready is a HuggingFace version of the IARPA Janus Benchmark-A (IJB-A) face recognition dataset in its CleanData layout, plus the official 10-split 1:1 verification and 1:N identification protocols. Unlike LFW-style benchmarks, IJB-A is template-based: each subject is represented by a template aggregating multiple still images and/or video frames. Protocol CSVs map every (template, file) row to a face annotation with bounding box, landmarks (e.g., right eye, left eye, nose base coordinates), and demographic attributes (e.g., gender, skin tone, age, facial hair). The dataset includes 5396 still images and 20369 video frames, totaling 25765 images, but protocols reference 25791 distinct files (with 26 video frames missing). The layout consists of image folders, a packed video frame archive, protocol CSV files (for training, verification, and identification), and baseline output files. Additionally, files.csv is provided for quick indexing and coverage checks. The frame.tar archive must be extracted to access video frames. The dataset is suitable for image classification and zero-shot image classification tasks, with a focus on face recognition, verification, and identification scenarios.