[1] Yanwei Fu, Timothy M. Hospedales, Tao Xiang, Jiechao Xiong, Shaogang Gong, Yizhou Wang, and Yuan Yao.  Robust Subjective Visual Property Prediction from Crowdsourced Pairwise Labels. IEEE TPAMI 2016

bib: @INPROCEEDINGS{ranking2016PAMI,
author = { Yanwei Fu and Timothy M. Hospedales and Tao Xiang and Jiechao Xiong and Shaogang Gong and Yizhou Wang and Yuan Yao},
title = {Robust Subjective Visual Property Prediction from Crowdsourced Pairwise Labels},
booktitle = {IEEE TPAMI},
year = {2016}

[2] Fu, Yanwei; Hospedales, T.; Xiang, T.; Gong, S; Yao. Y:Interestingness Prediction by Robust Learning to Rank, (ECCV 2014). Paper

bib: @INPROCEEDINGS{ranking2014ECCV,
author = { Yanwei Fu and Timothy M. Hospedales and Tao Xiang and Yuan Yao and Shaogang Gong},
title = {Interestingness Prediction by Robust Learning to Rank},
booktitle = {ECCV},
year = {2014}

[3] Ke Chen, Shaogang Gong, Tao Xiang, Chen Chang Loy, ``Cumulative attribute space for age and crowd density estimation,'' in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2013

To download all data:


The data explanations: (available if you start with http://www.eecs.qmul.ac.uk/~yf300/FG_NET_data/)
./images folder: all human face images. The groundtruth is used to name each image. For example, 078A11.JPG, means that this is the No.'78' person's image when he/she was 11 years old. 'A' is short for Age.

./points folder: this is the 68 manual annotated points for each image in ./images folder.
The annotated data is of much higher quality than another dataset e.g. MORPH (saved in /export/beware/thumper/yf300/Age_estimation_org_data_backup/ageEstimation/MOPRH). However, MORPH is much bigger dataset than FG-NET.

./feature_generation_tools: this is the tool to generate the features.
./feature_generation_tools/how-to-use-it: tutorial of how to use the tools.
./age50_10_round.mat is the 10 rounds of data used in my work [1].
Normally, you should firstly split the training/testing data by yourself. And generate the low-level feature for training/testing data respectively. For each split, the training/testing features are not the same. Because the process of generating training features is also needed to refer the annotations of testing features.

There is another very good tutorial and matlab labelling tool for AAM/ASM. You can download it from:
But some of them were written in Chinese.

PS: If any further questions, please email me: y.fu@qmul.ac.uk (and CC to ke.chen@tut.fi).
Yanwei Fu, Aug. 5th, 2014
Yanwei Fu, Aug. 22nd, 2014 --updated 


I am with the School of Data Science, Fudan University. I am working on computer vision : http://yanweifu.github.io/page3.html

Mar. 9th, 2022 -- updated