FLIC(Frames Labelled in Cinema)

FLIC 数据集包含来自好莱坞流行电影的 5003 张图像。这些图像是通过对 30 部电影的每十帧运行一个最先进的人体检测器获得的。被高度信任的人（大约 20,000 名候选人）随后被发送到众包市场 Amazon Mechanical Turk 以获得基本事实标签。每个图像都由 5 个 Turker 注释，以标记 10 个上身关节。在每幅图像中采用五个标记的中值，以对异常值注释具有鲁棒性。最后，如果该人被遮挡或严重非正面，则手动拒绝图像。

The FLIC dataset contains 5003 images sourced from popular Hollywood movies. These images were acquired by running a state-of-the-art human detector on every tenth frame across 30 films. A highly trusted pool of approximately 20,000 candidates were then assigned to the crowdsourcing marketplace Amazon Mechanical Turk to obtain ground-truth labels. Each image was annotated by 5 Turkers to mark 10 upper-body joints. The median of the five annotations for each image was employed to enhance robustness against outlier annotations. Finally, images where the depicted person was occluded or severely non-frontal were manually rejected.

搜集汇总

背景与挑战

背景概述

FLIC数据集是一个用于人体姿态估计的图像集合，包含5003张从好莱坞电影中提取的图像。这些图像通过人体检测器获取，并由众包标注了10个上身关节，标注过程采用5人标注取中值以增强鲁棒性，并手动排除了遮挡或非正面图像。