Context-aware Exposure Auto- correction. Global exposure auto-correction over-exposed under-exposed...

12
Context-aware Exposure Auto- correction

Transcript of Context-aware Exposure Auto- correction. Global exposure auto-correction over-exposed under-exposed...

Page 1: Context-aware Exposure Auto- correction. Global exposure auto-correction over-exposed under-exposed low-contrast input automatic histogram stretching.

Context-aware Exposure Auto-correction

Context-aware Exposure Auto-correction

Page 2: Context-aware Exposure Auto- correction. Global exposure auto-correction over-exposed under-exposed low-contrast input automatic histogram stretching.

Global exposure auto-correctionGlobal exposure auto-correction

over-exposed

under-exposed

low-contrast

input automatic histogram stretching

Page 3: Context-aware Exposure Auto- correction. Global exposure auto-correction over-exposed under-exposed low-contrast input automatic histogram stretching.

Global exposure auto-correctionGlobal exposure auto-correction

• Detection: valid histogram range < threshold

• Method: stretch histogram, adjust gamma curve

Test Images # Global Correction Percentage

1370 63 4.6%

#: Globally over-exposed, under-exposed & low-contrast images

Test Images include party, family, vacation, landscape, street view, pets

Page 4: Context-aware Exposure Auto- correction. Global exposure auto-correction over-exposed under-exposed low-contrast input automatic histogram stretching.

Local exposure auto-correctionLocal exposure auto-correction

input Local shadow / Highlight [ours]

• High dynamic range scene

Auto adjustment[WLPG]

Page 5: Context-aware Exposure Auto- correction. Global exposure auto-correction over-exposed under-exposed low-contrast input automatic histogram stretching.

Local exposure auto-correctionLocal exposure auto-correction

input Local shadow / Highlight [ours]

• Back-lighting object

Auto adjustment[WLPG]

Page 6: Context-aware Exposure Auto- correction. Global exposure auto-correction over-exposed under-exposed low-contrast input automatic histogram stretching.

High dynamic range scene detectionHigh dynamic range scene detection

input

confidence map of sky

sky detection

sky region scene region

segment

sky histogram scene histogram

local contrastin scene region

extract features

classifier

, ,

Page 7: Context-aware Exposure Auto- correction. Global exposure auto-correction over-exposed under-exposed low-contrast input automatic histogram stretching.

Samples of high dynamic range sceneSamples of high dynamic range scene

• True: • False:

Page 8: Context-aware Exposure Auto- correction. Global exposure auto-correction over-exposed under-exposed low-contrast input automatic histogram stretching.

High dynamic range sceneHigh dynamic range scene

True HDR scene # Percentage False HDR Scene

190 13.87% 126

#: True HDR scene images / Test Images

Test Images detected sky image Percentage

1370 316 23%

Page 9: Context-aware Exposure Auto- correction. Global exposure auto-correction over-exposed under-exposed low-contrast input automatic histogram stretching.

Back-lighting object detectionBack-lighting object detection

input

face detection

body detection

input

extract features

classifier

Histogram, local contrastin face/body

region

Histogram of image

• The most attractive backlit object is human!

Page 10: Context-aware Exposure Auto- correction. Global exposure auto-correction over-exposed under-exposed low-contrast input automatic histogram stretching.

Samples of back-lighting objectSamples of back-lighting object

• True: • False:

Page 11: Context-aware Exposure Auto- correction. Global exposure auto-correction over-exposed under-exposed low-contrast input automatic histogram stretching.

Back-lighting objectBack-lighting object

True backlit human # Percentage False backlit human

102 7.44% 493

#: True backlit human images / Test Images

Test Images

detected human image Percentage

1370 595 43.4%

Page 12: Context-aware Exposure Auto- correction. Global exposure auto-correction over-exposed under-exposed low-contrast input automatic histogram stretching.

SummarySummary

• Global incorrect exposure v.s. local incorrect exposure

• The “detected Human + Sky images” account for almost 66.5% of the whole test images

Global Local

4.6%21.31%

( hdr scene 13.87% + backlit human 7.44% )