salvacarrion/face-detection

一个用于训练经典人脸检测器（如Viola-Jones级联、Haar级联分类器或任何需要小灰度人脸裁剪和大量自然图像负样本的滑动窗口流程）的数据集。人脸裁剪来自四个知名源数据集，均预处理为灰度（单通道）并保持一致的方形裁剪。硬负样本源为Caltech-256，过滤掉了人脸/人物/人类类别，保留为原始彩色JPG，以便用户以任何分辨率提取负样本。数据集包含三个部分：训练集（约68k行，包含CelebA和FDDB的48×48人脸以及CBCL的19×19训练人脸/非人脸）、测试集（约24k行，包含CBCL的19×19测试人脸/非人脸）和负样本集（约30k行，包含过滤后的Caltech-256彩色JPG）。每行包含四列：图像（解码后的图像，CBCL/CelebA/FDDB为灰度，Caltech为彩色）、标签（0表示非人脸，1表示人脸）、来源（celeba、fddb、cbcl或caltech）和类别（Caltech类别文件夹名称，其他来源为null）。

A drop-in dataset for training classical face detectors (Viola-Jones-style cascades, Haar-cascade classifiers, or any sliding-window pipeline that needs small grayscale face crops + a large pool of natural-image negatives). The face crops come from four well-known source datasets, all preprocessed to grayscale (single channel) with consistent square cropping. The hard-negative source is Caltech-256 with face/people/human categories filtered out, kept as raw color JPGs so the user can extract negatives at any resolution. The dataset includes three splits: train (~68k rows, containing CelebA and FDDB 48×48 faces + CBCL 19×19 train faces/nofaces), test (~24k rows, containing CBCL 19×19 test faces + nofaces), and negatives (~30k rows, containing filtered Caltech-256 color JPGs). Each row has four columns: image (decoded image, grayscale for CBCL/CelebA/FDDB, color for Caltech), label (0 = noface, 1 = face), source ("celeba", "fddb", "cbcl", or "caltech"), and category (Caltech category folder name, null for other sources).