end to-end convolutional network for saliency prediction

28
End-to-end convolutional network for saliency prediction Junting Pan Xavier Giró-i-Nieto Slides online @DocXavi Large-scale Scene Understanding (LSUN) Challenge 2015 http://bit.ly/juntingnet

Upload: xavier-giro

Post on 05-Aug-2015

167 views

Category:

Technology


2 download

TRANSCRIPT

Page 1: End to-end convolutional network for saliency prediction

End-to-end convolutional network for saliency prediction

Junting Pan Xavier Giró-i-Nieto

Slides online@DocXavi

Large-scale Scene Understanding (LSUN)

Challenge 2015

http://bit.ly/juntingnet

Page 2: End to-end convolutional network for saliency prediction

2

Financial supportTechnical support

Albert Gil Josep Pujal

ACKNOWLEDGMENTS

Page 3: End to-end convolutional network for saliency prediction

3

LSUN SALIENCY CHALLENGE: A Déjà vu ?

John Markoff, “Scientists see promise in deep learning Programs”, The New York Times (Nov2012).

Photo: Keith Penner

Page 5: End to-end convolutional network for saliency prediction

5

LSUN SALIENCY CHALLENGE: A Déjà vu ?

Page 6: End to-end convolutional network for saliency prediction

6

RELATED WORK: Deep Saliency

Kümmerer, Matthias, Lucas Theis, and Matthias Bethge. "Deep Gaze I: Boosting Saliency Prediction with Feature Maps Trained on ImageNet." arXiv preprint arXiv:1411.1045 (2014).

Page 7: End to-end convolutional network for saliency prediction

7

RELATED WORK: Deep Saliency

Vig, Eleonora, Michael Dorr, and David Cox. "Large-scale optimization of hierarchical features for saliency prediction in natural images." Computer Vision and Pattern Recognition (CVPR), 2014 IEEE Conference on. IEEE, 2014.

Page 8: End to-end convolutional network for saliency prediction

8

RELATED WORK: Fully convolutional

Long, Jonathan, Evan Shelhamer, and Trevor Darrell. "Fully convolutional networks for semantic segmentation." Computer Vision and Pattern Recognition (CVPR), 2015 IEEE Conference on. IEEE, 2015.

Page 9: End to-end convolutional network for saliency prediction

9

RELATED WORK: Image Classification

CaffeNet

ARCHITECTURE[Khrizevsky’12]

DATA[Deng’09]

FRAMEWORK[Jia’14]

Page 10: End to-end convolutional network for saliency prediction

10

SALIENCY PREDICTION: JuntingNet

JuntingNet

Page 14: End to-end convolutional network for saliency prediction

14

SALIENCY PREDICTION: Architecture

Page 15: End to-end convolutional network for saliency prediction

15

SALIENCY PREDICTION: Architecture

End to end + regression = JuntingNet

Page 16: End to-end convolutional network for saliency prediction

16

SALIENCY PREDICTION: Architecture

Resize

96x96

Upsample + filter

4608 = 48x48

2D map

Page 18: End to-end convolutional network for saliency prediction

18

SALIENCY PREDICTION: Framework

Tutorial by Daniel Nouri (*) on regression for facial points for Kaggle.

(*) Daniel Nouri, “Using convolution networks to detect facil points” (Dec 2014).

on Lasagne

Page 19: End to-end convolutional network for saliency prediction

19

SALIENCY PREDICTION: Training

Data augmentation with horizontal mirroring.

Page 20: End to-end convolutional network for saliency prediction

20

SALIENCY PREDICTION: Training

Loss function Mean Square Error (MSE)

Weight initialization Gaussian distribution

Learning rate 0.03 to 0.0001

Mini batch size 128

Training time 7h (SALICON) / 3h (iSUN)

Acceleration Sigmoid + nesterov momentum 0.9

Regularisation Maxout norm

GPU NVidia GTX 980

Page 21: End to-end convolutional network for saliency prediction

21

RESULTS: Qualitative (iSUN)

JuntingNetGround TruthPixels

Page 22: End to-end convolutional network for saliency prediction

22

RESULTS: Qualitative (iSUN)

JuntingNetGround TruthPixels

Page 23: End to-end convolutional network for saliency prediction

23

RESULTS: Quantitative (iSUN)

Page 24: End to-end convolutional network for saliency prediction

24

RESULTS: Qualitative (SALICON)

JuntingNetGround TruthPixels

Page 25: End to-end convolutional network for saliency prediction

25

RESULTS: Qualitative (SALICON)

JuntingNetGround TruthPixels

Page 26: End to-end convolutional network for saliency prediction

26

RESULTS: Quantitative (SALICON)

Page 27: End to-end convolutional network for saliency prediction

27

RESULTS: Publications by end of June

http://bit.ly/juntingnet

Page 28: End to-end convolutional network for saliency prediction

28

Thank you LSUN ! Thank you Boston !

http://bit.ly/juntingnetSlides online @DocXavi