Subject Detection and Manipulation!
Eric Chang, Austin Lee and Samuel Yang!
contrast
Department of Electrical Engineering, Stanford University
Image Processing Pipeline
Motivation
Many interesting artistic manipulations can be done
on images and video if the subject can be reliably
detected:
•
•
•
Saliency maps
1
contrast
contrast
background blurring
automatic refocusing, exposure, white balance
object image search
center surround
center surround
2
color
average
original image
center
1
surround
center surround
average
finalsegmentation
segmentation map
content aware2
average
final segmentation map
Zheng, N., & Tang, X. (2007). Learning to Detect A Salient Object, 1–8. IEEE.
2 derived masks eveloped for the project
3 Goferman, S., Zelnik-Manor, L., & Tal, A. (2012). Context-aware saliency detection. IEEE PAMI.
desaturation map
final desaturated image
1
original
saliency map
Results
Using Jaccard similarity of the bounding boxes as
our performance metric, we increased accuracy to
68% over 200 images by using the three derived
saliency maps and our new segmentation scheme.
histogram count
content
color spatial distribution
color spatial distribution
3
aware
original image
contrast2
We then use the result to demonstrate an interesting
artistic manipulation, whereby we desaturate the
background and generate a simulated stereo view of
the subject, viewable as an animated GIF.
We used a subset of 200 images out of the 20,000
in [1], in which each image has exactly one salient
object, and for which a ground truth bounding box of
the salient object is known.
contrast
1
color
input
Here we implement and extend existing methods for
detecting salient objects.
Evaluation Metric
color spatial distribution
desaturation map
desaturated image
output