Saliency and Benchmarking

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
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]
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]
The progress?
• Salient objects in PASCAL
– 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).
The progress
FT: 0.28
GC: 0.39
SF: 0.35
PCA: 0.40
GC-seg: 0.38
The arguments
• No!!
These objects are not
• Our algorithm works
on images with salient
objects only!
The paradox of salient object
But hey, what is a
“salient object”?
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)
How to compose a biased salient
object detection dataset
Decide to build a new
salient object dataset!
Searching for unambiguous
examples of saliency…
Job done! Let other
people play with my
So what is saliency?
Found one! Add to
my dataset!
The Dataset Design bias
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]
• Dataset design bias: Biases introduced during
the design of a dataset:
– Exaggerating on stereotypical attributes.
• Limited variability in positive samples.
• Lack of negative samples at all.
Dataset design bias: the statistics
• Object number
Dataset design bias: the statistics
• Object eccentricity
Dataset design bias: the statistics
• Global foreground and background contrast
Dataset design bias: the statistics
• Local foreground/background contrast (contour strength)
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!
Data collection (in process)
• SR Research EyeLink 1000
• 2-sec viewing time.
• “Free-viewing” instruction (will mention it
• 3 subjects (more subjects on the way).
What makes an object salient
• Unit conversion:
– From fixation maps
– To object fixation score
• sum of blurred fixation map intensity within the object
Object size and saliency
• Large objects attract
more fixations.
• Small objects receive
denser fixations.
Object size and saliency
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)
Salient objects and salient object
• 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
Testing on salient objects
Salient objects on PASCAL VOC
FT: 0.22
GC: 0.35
SF: 0.31
PCA: 0.38
GC-seg: 0.39
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
Something is
The role of saliency in a visual system
• Bad performance because of boundary
• Bad performance because of unpredictability
of human “free will”?
Saliency as an oracle
• Oracle selecting the best segment
– CPMC: 78% from 154 segments
– gPB: 61% from 1286 segments
* coverage = intersect/union
Saliency and tasks
• Build a salient object detection dataset from
an egocentric object dataset.
• Let the eye-fixation speaks
Eye Tracker
Forward-looking Camera
Learning to recognize daily actions using gaze, [Fathi et. al. ECCV 12]
What makes an object salient?
• Object in egocentric actions
• Fixated object ==
Manipulated object?
• Joint work with Yin Li @
• Special thanks to Nathan
Faivre for his kind help on
eye tracking.
Open discussions

similar documents