HDR Reconstruction

1
HDR Reconstruction

• Across Temporally Varying Images

2
(No Transcript)
3
(No Transcript)
4
(No Transcript)
5
Quick Introduction

• Dynamic Range
• Ratio Brightest / Darkest
• Low Dynamic Range
• O(4001) Digital Camera
• High Dynamic Range
• O(10,0001) Human Eye

Data Courtesy dpreview.com ?Data Courtesy
Wikipedia ?Data Courtesy Gabriel Torres
(hardwaresecrets.com)
6
Low Dynamic Range (LDR)
?
Photo Courtesy of www.thepeaksislandhouse.com
7
High Dynamic Range
?
8
High Dynamic Range (HDR)

• Frequently Asked Questions (FAQs)
• Why is temporal variance bad?...

?
9
What caused those artifacts?
?
?
10
Why Was Response Curve _at_!?

• Response Curve
• Gives us irradiance (aka HDR data)
• If its ?, then our HDR image is ?
• Temporal Variance
• Baseline assumption in HDR Irradiance is
  constant in time!
• Jason screwed that up... ?
• We want a way to determine whether we can trust a
  given pixel to have constant irradiance

11
Recap Mathematics

• LDR ?
• HDR ?
• HDR(t) ?
• sHDR(t) ?

12
The first step towards getting somewhere is to
decide that you are not going to stay where you
are.
-Anonymous
13
The First Step

• End of the Day
• HDR image
• LDR regions
• Approach
• Separation of trusted/distrusted regions
• HDR recovery
• Stitching

14
Mosaic

• Rationale
• For a given pixel, ? a best exposure in the stack
• If we dont trust a pixel, we would like to at
  least use its best exposure
• Approach
• Pick common good exposure as base (1/40 sec)
• Find the ? regions (temporal variance)
• Compute their best exposure and store in mosaic
• Do HDR in ? regions
• Everything else replaced with mosaic

15
Example
16
Regions of Difference

• Definition
• Irradiance of adjacent images is not equal (?)
• Acquisition
• Generate cameras response curve
• Use assumption that irradiance is constant in
  time
• Group pixels together

17
(No Transcript)
18
Jason_Poltergeist.jpg... (Seriously)
19
(No Transcript)
20
(No Transcript)
21
(No Transcript)
22
(No Transcript)
23
(No Transcript)
24
Actual Methodology

• Construct Response
• Initial estimate, cut off ends (manually)
• Use to calculate ln(E)
• Find RODs
• Initialize points based on ROD(z) equation
• Expand points by kernel
• Throw away solitary points
• What Next?...

25
Graph

• Nodes
• Each ROD becomes a node in the graph
• Connect adjacent nodes
• Cut RODs based on weights (yields best exposed
  ROD)
• This becomes new image, find RODs with next
  image, repeat
• Weights
• Each node receives weight

26
Recovering the Mosaic
?
?
27
sHDR Reconstruction

• Trusted/Distrusted Pixels
• Was pixel i ever in an ROD?
• Yes Distrusted ?
• No Trusted ?
• Now, run HDR Reconstruction on Trusted Pixels
• Anything else is just the mosaic
• End Product
• HDR in Trusted Region
• Mosaic in Distrusted Region

28
(No Transcript)
29
Jason Takes On...
30
Wrap-Up

• The Future of sHDR
• Graph cuts need to be non-human
• Dithering between HDR and LDR
• Acknowledgements
• Jason Lawrence
• Uyttendale et. al
• Debevec et. al
• Jiajun
• Rockstar Energy Drink

31
Accidental Art
32
I can't change the direction of the wind, but I
can adjust my sails to always reach my
destination.
-Jimmy Dean
33
y a-t-il questions?