gaobb / DLDL

[TIP] Deep Label Distribution Learning with Label Ambiguity

DLDL-MatConvNet

This repository is a MatConvNet re-implementation of "Deep Label Distribution Learning with Label Ambiguity", Bin-Bin Gao, Chao Xing, Chen-Wei Xie, Jianxin Wu, Xin Geng. The paper is accepted at [IEEE Trans. Image Processing (TIP), 2017].

You can train Deep ConvNets from Scratch or a pre-trained model on your datasets with limited samples and ambiguous labels. This repo is created by Bin-Bin Gao.

Table of Contents

0. Facial Age Estimation
1. Head Pose Estimation
2. Multi-label Classification
3. Semantic Segmentation

Facial Age Estimation

step1: download pre-trained model to ./DLDLModel

step2: in matlab, run age-demo.m

Pre-trained models:

Dataset	Model	MAE	epsilon-error
ChaLearn15	DLDL	5.34(exp)	0.44
ChaLearn15	DLDL+VGG-Face	3.51(exp)	0.31
Morph	DLDL	2.51±0.03 (max)	-
Morph	DLDL+VGG-Face	2.42±0.01 (max)	-

Head Pose Estimation

step1: download pre-trained model to ./DLDLModel

step2: in matlab, run pose-demo.m

Pre-trained models:

Dataset	Model	Pitch	Yaw	Pitch+Yaw	Pitch	Yaw	Pitch+Yaw
Pointing’04	DLDL	1.69±0.32	3.16±0.07	4.64±0.24	91.65±1.13	79.57±0.57	73.15±0.72
BJUT-3D	DLDL	0.02±0.01	0.07±0.01	0.09±0.01	99.81±0.04	99.27±0.08	99.09±0.09
AFLW	DLDL	5.75	6.60	9.78	95.41	92.89	89.27

Multi-label Classification

step1: download pre-trained model to ./DLDLModel

step2: in matlab, run ml-demo.m

Single-model classification resluts (mAP in %) on VOC2007

Training Style	Net-D+Max	Net-D+Avg	Net-E+Max	Net-E+Avg
IF-DLDL	90.1 model	90.5 model	90.6 model	90.7 model
PF-DLDL	92.3 model	92.1 model	92.5 model	92.2 model

Multi-model ensemble results (mAP in %) on VOC2007 and VOC2012

Dataset	Training Style	mAP
VOC2007	IF-DLDL	91.1
VOC2007	PF-DLDL	93.4
VOC2012	IF-DLDL	89.9
VOC2012	PF-DLDL	92.4

Semantic Segmentation

step1: download pre-trained model to ./DLDLModel

step2: in matlab, run seg-demo.m

Dataset	Model	MIoU
VOC2011	DLDL-8s	64.9
VOC2011	DLDL-8s+CRF	67.6
VOC2012	DLDL-8s	64.5
VOC2012	DLDL-8S+CRF	67.1

Additional Information

If you find DLDL helpful, please cite it as

@ARTICLE{gao2016deep,
         author={Gao, Bin-Bin and Xing, Chao and Xie, Chen-Wei and Wu, Jianxin and Geng, Xin},
         title={Deep Label Distribution Learning with Label Ambiguity},
         journal={IEEE Transactions on Image Processing},
         year={2017},
         volume={26},
         number={6},
         pages={2825-2838}, 
         }

ATTN1: This packages are free for academic usage. You can run them at your own risk. For other purposes, please contact Prof. Jianxin Wu ([e-mail]).