Daimler Monocular Pedestrian Detection

戴姆勒单目行人检测涉及大量训练和测试集。训练集包含 15.560 个行人样本（48x96 分辨率的图像剪切）和 6744 个额外的不包含行人的完整图像，用于提取负样本。测试集包含一个独立的序列，其中包含超过 21.790 个图像和 56.492 个行人标签（完全可见或部分遮挡），在 27 分钟车程通过城市交通的车辆中以 VGA 分辨率（640x480，未压缩）捕获。因此，该数据集是真实的，并且在发布时比其他数据集大一个数量级（8.5 Gb）。

The Daimler Monocular Pedestrian Detection dataset includes large-scale training and test sets. The training set contains 15,560 pedestrian samples (cropped images with a resolution of 48×96) and 6,744 additional full-frame images without any pedestrians for extracting negative samples. The test set consists of an independent sequence with over 21,790 images and 56,492 pedestrian annotations (either fully visible or partially occluded), which was captured at VGA resolution (640×480, uncompressed) from a vehicle traveling through urban traffic during a 27-minute drive. Therefore, this is a real-world dataset that was an order of magnitude larger than other contemporary datasets upon its release, with a total size of 8.5 Gb.

搜集汇总

背景与挑战

背景概述

该数据集由海德堡大学和阿姆斯特丹大学于2009年发布，专注于单目行人检测。训练集包含15,560个行人样本和6,744个无行人图像，测试集则提供21,790张VGA分辨率图像和56,492个行人标签，数据源自城市交通场景的车辆捕获，总大小约8.5GB，在发布时规模领先于同类数据集。