发新帖

WLFDB人脸数据库下载,WLFDB人脸数据库介绍

  [复制链接]
5002 189
WLFDB人脸数据库介绍:
QQ截图20151231121547.jpg

Auto face annotation aims to detect human faces from a photo image and tag the facial image with the human names, is a fundamental research problem and beneficial to many real-world applications. In general, there are two ways to tackle the automatic annotation problem, one is the model-based face annotation scheme, which is formulated as a traditional face recognition problem; another is the data-driven based face annotation, which recently has been attracting more and more attentions to handle the large mount of freely available data over the internet.
Recently, several web facial image databases are constructed for face verification or weakly supervised model learning. For example, LFW, Yahoo!News and FAN-Large are constructed based on new images and the corresponding caption texts. Nevertheless, these three databases are not suitable for data-driven based face annotation, since the number of images for each person is quite small. Moreover, Pubfig dataset is an expedient choice, which contains 200 persons and abut 50 thousand images. Nevertheless, the number of subjects is a bit small to evaluate large-scale data-driven based face annotation techniques. In light of this situation, we release a well-prepared large-scale web facial images database: Weakly Labeled Faces on the web (WLFDB) for the data-driven based face annotation problem database to facilitate the related research works.
The WLFDB dataset is constructed following three principles:
Large-scale Weakly Labeled Images
We aim to build a large-scale image dataset for the data-driven scheme. Consequently, WLFDB contains 6,025 subjects and over 0.7 million web facial images. For each subject, there are over 100 images on average. Different with the existing datasets, the labels of WLFDB are automatically assigned based on the text query, which make it extremely easy to be extended without extra manual efforts.
Comprehensive Data Definition
We aim to make a comprehensive data deification in WLFDB, which contains three level data types: "raw web facial images", "aligned facial images", and "facial feature representation". For each data type, we make WLFDB easy to be used without exploring complex extra processing steps. For example, the annotation technique can be directly evaluated by using the provide three kinds of feature representation.
Standard Evaluation Portal
We provide a standard evaluation portal to evaluate the performance of data-driven based face annotation scheme. In particular, we build a manually labeled ground-truth query set, which contains 119 peoples and 1,600 images. The query set is randomly divided into "training" and "testing" sets of equal size for 10 times. The "training" set is used to learn annotation models and tune parameters, while the "testing" set is used to evaluate the annotation performance. In addition, three baseline algorithms are ev aluated and compared over the aforementioned 10 "test" according to the hit rate performance metric.
In summary, WLFDB dataset is a large-scale weakly labeled web facial images database, which is built according to the "real" web facial images distributions. We hope it will facilitate the following research works.
包含了两个数据库,分别为DB0400和DB6000,其中DB0400包含了400个人,总共53448张人脸图片,而DB6000包含了6025个人,共714454张人脸图片。
DB0400数据库下载地址:
游客,如果您要查看本帖隐藏内容请回复

DB6000数据库下载地址:
游客,如果您要查看本帖隐藏内容请回复

更多的标签下载见:
游客,如果您要查看本帖隐藏内容请回复


举报 使用道具

回复

精彩评论189

Ben  新手上路  发表于 2015-12-31 13:16:26 | 显示全部楼层
有没有针对新疆或者中亚地区的人脸库?

举报 使用道具

回复 支持 反对
echofeng  中级会员  发表于 2015-12-31 14:15:17 | 显示全部楼层
多谢分享~~~

举报 使用道具

回复
来者不拒  注册会员  发表于 2015-12-31 14:25:58 | 显示全部楼层
感谢分享

举报 使用道具

回复
bowrain  中级会员  发表于 2015-12-31 23:20:16 | 显示全部楼层
好大的库

举报 使用道具

回复
NotThere  注册会员  发表于 2016-1-1 09:46:01 | 显示全部楼层
确实每人图片太少不足以model类内的variance。另外的问题是groundtruth标注 希望比较准

举报 使用道具

回复 支持 反对
chenriwei  版主  发表于 2016-1-1 22:54:47 | 显示全部楼层
这年头,缺的就是数据了。赞

举报 使用道具

回复 支持 反对
chenriwei  版主  发表于 2016-1-1 23:01:21 | 显示全部楼层
We use OpenCV for face detection, and adopt the Deformable Lucas-Kanade ( DLK ) algorithm for face alignment. The non-face-detected web images are ignored directly.


举报 使用道具

回复 支持 反对
frostjohn  中级会员  发表于 2016-1-2 20:28:04 | 显示全部楼层
多谢分享

举报 使用道具

回复
hzwhhh  新手上路  发表于 2016-1-3 18:14:57 | 显示全部楼层
谢谢楼主分享

举报 使用道具

回复 支持 反对
*滑动验证:
您需要登录后才可以回帖 登录 | 立即注册

本版积分规则

更多

关注我们

QQ:448109455 周一至周日8:30-20:30
快速回复 返回顶部 返回列表