Tutorial Overview - MIT Saliency...

New directions in saliency research: Developments in architectures, datasets, and evaluation

Tutorial OverviewZoya Bylinskii & Tilke Judd, ECCV Oct. 8, 2016

Ali BorjiCenter for Research in

Computer Vision,University of Central Florida

Zoya BylinskiiComputer Science and

Artificial Intelligence Lab,MIT

Tilke JuddProduct Manager,

GooglePreviously: MIT

MIT Saliency Benchmark Team

Tutorial organized by:

Judd et al. “A Benchmark of Computational Models of Saliency to Predict Human Fixations” [MIT Tech Report 2012]

• 10 saliency models

• 3 baslines

• 3 metrics (AUC, SIM, EMD)

• 1 dataset

• 66 saliency models

• 5 baslines

• 8 metrics

• 2 datasets

2011 2016

large boosts in performance

in recent yearsdriven by neural

network architectures

tracking progress with the MIT Saliency Benchmark

• 66 models evaluated on MIT benchmark (as of Sept 30, 2016)

• top 8/10 are neural networks

• neural networks make up 25% of all submissions (17 models since 2014)

What are the common architectures used for neural network models of saliency?

How are these models trained?

PDP - Jetley et al. [CVPR 2016]

DeepGaze - Kümmerer et al. [ICLR 2015 workshop]

ML-Net - Cornia et al. [ICPR 2016]

SALICON - Huang et al. [ICCV 2015]

DeepFix - Kruthiventi et al. [ArXiv 2015]

eDN - Vig et al. [CVPR 2014]

loss functions & evaluation

datasets

architectures

applications

semanticsegmentation

objectdetection

objectproposals

image clustering & retrieval

cognitivesaliency

Tutorial overview“Deep networks for

saliency map prediction” Naila Murray

tracking progress with the MIT Saliency Benchmark

• 66 models evaluated on MIT benchmark

• using 8 evaluation metrics

• on 2 datasets (MIT300, CAT2000)

tracking progress with the MIT Saliency Benchmark

0 0.2 0.4 0.6 0.8 1159

1317212529333741454953576165

AUC-Judd

0 0.2 0.4 0.6 0.8159

1317212529333741454953576165

sAUC

0 0.5 1 1.5 2 2.5159

1317212529333741454953576165

NSS

0 0.2 0.4 0.6 0.8 11591317212529333741454953576165

CC

0 0.2 0.4 0.6 0.8159

1317212529333741454953576165

SIM

gradual improvementover time

but saturating

tracking progress with the MIT Saliency Benchmarkm

odel

s

0 0.2 0.4 0.6 0.8 1159

1317212529333741454953576165

AUC-Judd

0 0.2 0.4 0.6 0.8159

1317212529333741454953576165

sAUC

0 0.5 1 1.5 2 2.5159

1317212529333741454953576165

NSS

0 0.2 0.4 0.6 0.8 11591317212529333741454953576165

CC

0 0.2 0.4 0.6 0.8159

1317212529333741454953576165

SIM

tracking progress with the MIT Saliency Benchmarkm

odel

s

0 0.2 0.4 0.6 0.8 1159

1317212529333741454953576165

AUC-Judd

0 0.2 0.4 0.6 0.8159

1317212529333741454953576165

sAUC

0 0.5 1 1.5 2 2.5159

1317212529333741454953576165

NSS

0 0.2 0.4 0.6 0.8 11591317212529333741454953576165

CC

0 0.2 0.4 0.6 0.8159

1317212529333741454953576165

SIM

large boosts in performance

in recent yearsdriven by neural

network architectures

tracking progress with the MIT Saliency Benchmarkm

odel

s

As model performances continue to approach ground truth, how can we have

more meaningful evaluation?

loss functions & evaluation

datasets

architectures

applications

semanticsegmentation

objectdetection

objectproposals

image clustering & retrieval

cognitivesaliency

Tutorial overview“Evaluating saliency models in a

probabilistic framework”Matthias Kümmerer

New data collection methodologies

How can we collect enough training data for neural network models of saliency?

Krafka et al. [CVPR 2016]

Papoutsaki et al. [IJCAI 2016]

Xu et al. [ArXiv 2015]

New data collection methodologies

iSUN

Webcam-based

Krafka et al. [CVPR 2016]

Papoutsaki et al. [IJCAI 2016]

Xu et al. [ArXiv 2015]

New data collection methodologies

Kim et al. [CHI EA 2015]

Jiang et al. [CVPR 2015]

iSUN SALICON

Click-basedWebcam-based

What kind of new applications are possiblewith state-of-the-art saliency architectures

and big data?

loss functions & evaluation

datasets

architectures

applications

semanticsegmentation

objectdetection

objectproposals

image clustering & retrieval

cognitivesaliency

Tutorial overview“Saliency for image

understanding and manipulation”Ming-Ming Cheng

Is saliency solved?

Is saliency solved?NO!

“Towards cognitive saliency”Zoya Bylinskii

Tutorial overview

Gaze Text Face Action Main Focus

Inpu

tH

uman

Fi

xatio

nsM

odel

Pr

edic

tions

“Towards cognitive saliency”Zoya Bylinskii

Tutorial overview