model comparison and challenges ii compositional bias of salient object detection benchmarking

35
Model comparison and challenges II Compositional bias of salient object detection benchmarking Xiaodi Hou K-Lab, Computation and Neural Systems California Institute of Technology for the Crash Course on Visual Saliency Modeling: Behavioral Findings and Computational Models CVPR 2013

Upload: ugo

Post on 25-Feb-2016

39 views

Category:

Documents


0 download

DESCRIPTION

Model comparison and challenges II Compositional bias of salient object detection benchmarking. for the Crash Course on Visual Saliency Modeling: Behavioral Findings and Computational Models CVPR 2013. Xiaodi Hou K-Lab, Computation and Neural Systems California Institute of Technology. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Model comparison and challenges II Compositional bias of salient object detection benchmarking

Model comparison and challenges IICompositional bias of salient object detection benchmarking

Xiaodi HouK-Lab, Computation and Neural Systems

California Institute of Technology

for the Crash Course on Visual Saliency Modeling:Behavioral Findings and Computational Models

CVPR 2013

Page 2: Model comparison and challenges II Compositional bias of salient object detection benchmarking

Schedule

Page 3: Model comparison and challenges II Compositional bias of salient object detection benchmarking

On detecting salient objects• Learning to Detect A Salient Object [Liu et. al., CVPR 07]• Frequency-tuned Salient Region Detection [Achanta et. al., CVPR 09]

Page 4: Model comparison and challenges II Compositional bias of salient object detection benchmarking

The progress!• Some top performers:

– [PCA] What makes a patch distinct [Margolin et. al., CVPR 13]– [SF]Saliency filters [Perazzi et. al., CVPR 12]:

• F-Measure: 0.84– [GC]/[GC-seg]Global contrast-based salient region detection [Cheng et. al., CVPR 11]

• F-Measure: 0.75– [FT] Frequency Tuned Salient Region Detection [Achanta et. a.l., CVPR 09] :

• 0.65 by [Achanta et. al., CVPR 09].

Image from [Perazzi et. al., CVPR 2012]

Page 5: Model comparison and challenges II Compositional bias of salient object detection benchmarking

The progress?• Salient objects in PASCAL

VOC?– 850 images from VOC 2013

validation set.– Intersection of main challenge

and segmentation challenge.– Answers more questions:• Where is your algorithm (in

salient object detection)?• Where is salient object detection

(in computer vision).

Page 6: Model comparison and challenges II Compositional bias of salient object detection benchmarking

The progress

• FT: 0.28• GC: 0.39• SF: 0.35• PCA: 0.40• GC-seg: 0.38

55% performance

drop!!

Page 7: Model comparison and challenges II Compositional bias of salient object detection benchmarking

The arguments

• No!!These objects are not salient!

• Our algorithm works on images with salient objects only!

Page 8: Model comparison and challenges II Compositional bias of salient object detection benchmarking

The paradox of salient object detection

But hey, what is a “salient object”?

Page 9: Model comparison and challenges II Compositional bias of salient object detection benchmarking

COMPOSITIONAL BIAS

Page 10: Model comparison and challenges II Compositional bias of salient object detection benchmarking

Before we proceed…

• Google Image Search: “science”– Rutherford atomic model (9)– Test tubes (10)– Microscopes (4)– Double helix (3)– Old guys with crazy hair and glasses (3)

Stereotypes of science are not sciences!

Page 11: Model comparison and challenges II Compositional bias of salient object detection benchmarking

How to compose a biased salient object detection dataset

Decide to build a new salient object dataset!

So what is saliency?

Searching for unambiguous examples of saliency…

Found one! Add to my dataset!

Job done! Let other people play with my

dataset!

Page 12: Model comparison and challenges II Compositional bias of salient object detection benchmarking

The compositional bias

• Compositional bias: Biases introduced during the composition of a dataset:– Exaggerating on stereotypical attributes.• Limited variability in positive samples.• Lack of negative samples at all.

Unlike datasets in machine learning, where the dataset is the world, computer vision datasets are supposed to be a representation of the world.

---- [Torralba and Efros: Unbiased look at Dataset bias]

Page 13: Model comparison and challenges II Compositional bias of salient object detection benchmarking

Compositional bias: the statistics

• Object number

Page 14: Model comparison and challenges II Compositional bias of salient object detection benchmarking

Compositional bias: the statistics

• Object eccentricity

Page 15: Model comparison and challenges II Compositional bias of salient object detection benchmarking

Compositional bias: the statistics• Global foreground and background contrast

Page 16: Model comparison and challenges II Compositional bias of salient object detection benchmarking

Compositional bias: the statistics• Local foreground/background contrast (contour strength)

Page 17: Model comparison and challenges II Compositional bias of salient object detection benchmarking

TOWARDS A BETTER SALIENT OBJECT DATASET

Page 18: Model comparison and challenges II Compositional bias of salient object detection benchmarking

The new project

• Build a salient object detection dataset from a good object detection dataset (e.g. PASCAL VOC).

Let the eye fixations pick up those salient objects!

Page 19: Model comparison and challenges II Compositional bias of salient object detection benchmarking

Data collection (in process)

• SR Research EyeLink 1000• 2-sec viewing time.• “Free-viewing” instruction (will mention it later).• 3 subjects (more subjects on the way).

We will release the dataset very soon!

Page 20: Model comparison and challenges II Compositional bias of salient object detection benchmarking

What makes an object salient• Unit conversion:– From fixation maps– To object fixation score• sum of blurred fixation map intensity within the object

mask.

Page 21: Model comparison and challenges II Compositional bias of salient object detection benchmarking

Object size and saliency

• Large objects attract more fixations.

• Small objects receive denser fixations.

Page 22: Model comparison and challenges II Compositional bias of salient object detection benchmarking

Object size and saliency

Page 23: Model comparison and challenges II Compositional bias of salient object detection benchmarking

Objects, salient objects, and the most salient objects

• Salient objects:– Fixation score higher than

mean (67.3% objects).• Most salient objects:– Fixation score higher than

mean*2 (27.8% objects).

Image with fixation Object labeling Salient objects Most salient object(s)

Page 24: Model comparison and challenges II Compositional bias of salient object detection benchmarking

Salient objects and salient object detection

• Guess how does the algorithms perform on “salient objects” and “most salient objects”?

On all objects:• FT: 0.28• GC: 0.39• SF: 0.35• PC: 0.38

Page 25: Model comparison and challenges II Compositional bias of salient object detection benchmarking

Testing on salient objects

Salient objects on PASCAL VOC

60% performance

drop!!

• FT: 0.22• GC: 0.35• SF: 0.31• PCA: 0.38• GC-seg: 0.39

Page 26: Model comparison and challenges II Compositional bias of salient object detection benchmarking

Testing on most salient objects

Most salient objects on PASCAL VOC

• FT: 0.10• GC: 0.20• SF: 0.15• PCA: 0.26• GC-seg: 0.23

79.8% performance

drop!!

Page 27: Model comparison and challenges II Compositional bias of salient object detection benchmarking

Something is wrong, seriously!

Page 28: Model comparison and challenges II Compositional bias of salient object detection benchmarking

DISCUSSIONS

Page 29: Model comparison and challenges II Compositional bias of salient object detection benchmarking

The role of saliency in a visual system

• Bad performance because of boundary detection?

• Bad performance because of unpredictability of human “free will”?

Page 30: Model comparison and challenges II Compositional bias of salient object detection benchmarking

Saliency as an oracle

• Oracle selecting the best segment– CPMC: 78% from 154 segments– gPB: 61% from 1286 segments

* coverage = intersect/union

Page 31: Model comparison and challenges II Compositional bias of salient object detection benchmarking

Saliency and tasks

• Build a salient object detection dataset from an egocentric object dataset.

• Let the eye-fixation speaks

Eye TrackerForward-looking Camera

Learning to recognize daily actions using gaze, [Fathi et. al. ECCV 12]

Page 32: Model comparison and challenges II Compositional bias of salient object detection benchmarking

What makes an object salient?Task

ObjectSaliency

• Object in egocentric actions• Fixated object ==

Manipulated object?

Page 33: Model comparison and challenges II Compositional bias of salient object detection benchmarking

THANKS

Page 34: Model comparison and challenges II Compositional bias of salient object detection benchmarking

Acknowledgement

• Joint work with Yin Li @ Gatech.

• Special thanks to Nathan Faivre for his kind help on eye tracking.

Page 35: Model comparison and challenges II Compositional bias of salient object detection benchmarking

Open discussions