Burst Photography - Stanford Universityweb.stanford.edu/class/ee367/slides/lecture7.pdf ·...
Transcript of Burst Photography - Stanford Universityweb.stanford.edu/class/ee367/slides/lecture7.pdf ·...
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
●
●
●
●
•
• sparse matrix
•
•
•
•
•
•
•
• →
→
•
•
•
•
•
•
4.7 mm2
canon.com
1360 mm24.7 mm2
Larger aperture Long exposure Flash
How to collect more light? Capture multiple photos
video denoising[Kokaram, 1993]
[Dabov et al., 2007][Liu et al., 2014]
exposure bracketing[Debevec et al., 2007][Gallo and Sen 2016]
image stacking
weatherandsky.com
1 frame
16 frames
[Fruchter and Hook, 2002]
● burst with constant exposure○ more robust merge○ underexposed up to 8x
● reference image○ physically consistent fallback
● raw images○ merge in raw too
● burst with constant exposure○ more robust merge○ underexposed up to 8x
● burst with constant exposure○ more robust merge○ underexposed up to 8x
● reference image○ physically consistent fallback
Our approach
exposure bracketing● higher SNR● challenging merge
● HDR capture as noise reduction [Hasinoff et al. 2010] [Zhang et al. 2010]
Underexposure for HDR
single underexposed shot● low SNR
underexposed burst● moderate SNR● more robust merge ...
...
12 MPix, 10-bit
System overview
customautoexposure
processing
align & merge
12 MPix, 14-bit
zeroshutterlagcapture
finish
12 MPixring
buffer
Autoexposure and tonemapping
Example-based autoexposure
AE database~5,000 scenes
scenedescriptor
AE result
scenedescriptor
dual AE result - controls capture and rendition
shortexposure
longexposure
AE result
raw input
● synthetic exposures from AE○ single merged input image○ digital gains
● automatically set tuning parameters
Local tonemapping
pyramid blending for HDR tonemapping[Mertens et al., 2007]
[Paris et al. 2011][Aubry et al. 2014]
Align and merge
Burst alignment
reference frame
R+Gr+Gb+B
input burst
local alignmentto reference
reference frame
alternate frame
raw
120x90 (or 60x45)12 MPix3 MPix
Coarse to fine alignment
4x
4x
● 4 pyramid levels● upsample with multiple hypotheses [Tao et al., 2012]
metric search range subpixel?
L1
L2 9x9
3x3 no
yes
2x
Example alignment
aligned to reference
reference frame
Tiled Fourier-based merge● divide into 16x16 or 32x32 tiles
○ 50% overlap - every pixel covered by 4 tiles● merge in Fourier domain
...
reference tile
aligned alternate tiles
merged tile
Robust per-frequency merge
robust pairwise merge
reference frame
aligned average
Finish
Finish pipeline
demosaic,white balance
chroma denoise
lens shading, color correction
local tonemapping
global tonemapping
sharpen
suppress aberrations
color tuning YUV→U'V' LUT
merged raw
final result
12 MPix, 14-bit
12 MPix, 8-bit
auto exposure
luma
chroma
Results
HDR+ Off
HDR+ On
HDR+ Off
HDR+ On
Released Nov. 2018
Siggraph Asia 2019
HDR+ Night Sight
Night Sight● Dedicated mode
○ positive shutter lag○ heavier processing
● More light○ bigger bursts○ much longer exposures
● Motion metering● Stronger merge
○ more aggressive○ super-resolution merge
● ML-based AWB● Night "look"
○ brighter AE○ more HDR
● less motion?○ longer exposure○ better SNR
Measure scene motion in real time, to choose the exposure time
Motion metering
motion metering off8.3 ms, ISO 346
motion metering on59 ms, ISO 53
● more motion?○ shorter exposure○ less motion blur
● less motion?○ longer exposure○ better SNR
Measure scene motion in real time, to choose the exposure time
Motion metering
● more motion?○ shorter exposure○ less motion blur
motion metering off17 ms, ISO 183
motion metering on6 ms, ISO 512
Motion metering - "tripod detection"
● When camera is very stable, allow even longer exposures● Combine signals from gyro (camera motion) and images
(camera+scene motion)○ Discount motion due to shutter tap
● For fixed time budget, prefer fewer frames○ Less read noise start of burst
77
Without tripod detection 15 frames x 333ms
With tripod detection 6 frames x 1000ms
78
Without tripod detection 15 frames x 333ms
With tripod detection 6 frames x 1000ms
Challenges for Night Sight merge
object motion camera motion low SNR
Merge increases →High noise, motion robust Low noise, motion blur
80
Motion-Adaptive Burst Merging
Modifying the merge strength
HDR+ Increased merge
HDR+Temporal strength = 16
Increased mergeTemporal strength = 200
Ours: Tile-wise merge using mismatch maps
Modified merge: spatially varying merge strength
Super Res Zoom on Pixel 3
Siggraph 2019
White balance gains are applied to make “white” appear white
Captured image
Predicted “illumination”
Predicted imageInverse rgb gains
=/
White balance is an ill-posed problem
Learning based white balance: "Fast Fourier Color
Constancy", Barron & Tsai, 2017
Extended to low-light scenes
Challenges of Color Constancy in low light
Noisier images
➡ Train on real images that have noise
➡ New training data set for low light (3,500 images)
Various illuminants:Highly colorful illuminants result in color channels with practically zero signal (“missing channels”)
➡ New error metric: Anisotropic Reproduction Error (ARE)
Auto white balance results
Left: camera default, Right: ours
Baseline system(HDR+, Hasinoff et
al, 2016)
13 frames,exposure time:
1/15 s
Adding new tone mapping to
baseline
Results are noisy
Adding motion metering
Exposure time 1/3 s
Noise and detail improve
Adding motion-robust
merge
Noise and detail improve further
Adding motion-robust
merge
Noise and detail improve further
Adding low-light auto white
balance
Color improves
Side by side comparison
Previously described result (HDR+, Hasinoff et al 2016) Our result
Results across a variety of scenes: indoor
Previously described result (HDR+, Hasinoff et al 2016) Our result
Results across a variety of scenes: outdoor
Previously described result (HDR+, Hasinoff et al 2016) Our result
“Learning to see in the dark”, Chen et al, 2018
End to end network
Training set is pairs on short and long exposures
Comparison to “Learning to See in the Dark” (Chen et al, 2018)
Chen et al 2018 Our result
Comparison to “Learning to See in the Dark” (Chen et al, 2018)
Chen et al 2018 Our result
Astrophotography
Astrophotography
Extending exposure time
● Up to 4 minutes total exposure: ○ Capture even more light to improve signal to noise○ Extend the capture time per frame
● Detect when the camera is static using gyro signals● Detect when the scene is static
● Traditional: extend the exposure time● Computational:
● Burst photography● Motion metering● Motion-adaptive merge● Tone mapping for nighttime scenes● Device stability detection● “Warm pixel” removal
● Machine learning:● Learning based white balance● Semantically aware image processing of the skies
● Implemented on a mobile device
In Summary,To Capture Low Light Scenes:
2019 Smartphone Camera of the year award“If you shoot Night Sight - even during daylight hours - you'll be rewarded with some of the best detail retention and balanced noise reduction we've seen from a smartphone... A new astrophotography mode is not just cool but inspiring, and also benefits any nighttime scene where longer exposures can be used. The combination of super-res zoom and a new telephoto module make 'zoomed in' photos better than many peers.” dpreview.com
HDR
• Mann, Picard “On Being ‘Undigital’ with Digital Cameras: Extending Dynamic Range by Combining Differently Exposed Pictures”, IS&T 1995
• Debevec, Malik, “Recovering High Dynamic Range Radiance Maps from Photographs”, SIGGRAPH 1997
• Robertson, Borman, Stevenson, “Estimation-Theoretic approach to Dynamic Range Improvement Using Multiple Exposures”, Journal of Electronic Imaging 2003
• Mitsunaga, Nayar, “Radiometric self Calibration”, CVPR 1999
• Reinhard, Ward, Pattanaik, Debevec (2005). High dynamic range imaging: acquisition, display, and image-based lighting. Elsevier/Morgan Kaufmann
• Fattal, Lischinski, Werman, “Gradient Domain High Dynamic Range Compression”, ACM SIGGRAPH 2002
• Durand, Dorsey, “Fast Bilateral Filtering for the Display of High Dynamic Range Images”, ACM SIGGRAPH 2002
Super-resolution
• Baker, Kanade, Limits on super-resolution and how to break them“ IEEE Transactions on Pattern Analysis and Machine Intelligence 24(9), 1167–1183 (2002)
• Ben-Ezra, Lin, Wilburn, Zhang,, “Penrose pixels for super-resolution” EEE Transactions on Pattern Analysis and Machine Intelligence 33(7), 1370–1383 (2011)
• Ben-Ezra, Zomet, Nayar, “Jitter Camera: High Resolution Video from a Low Resolution Detector”, CVPR 2004
• Ben-Ezra, Zomet, Nayar, “Video super-resolution using controlled subpixel detector shifts” IEEE Trans. PAMI27(6), 977–987 (2005)
• Elad, Feuer, “Restoration of single super-resolution image from several blurred, noisy and down-sampled measured images” IEEE Trans. Im. Proc. 6(12), (1997)
Other
• Hasinoff, Sharlet, Geiss, Adams, Barron, Kainz, Chen, Levoy “Burst photography for high dynamic range and low-light imaging on mobile cameras”, SIGGRAPH Asia 2016
• Orly Liba, Kiran Murthy, Yun-Ta Tsai, Tim Brooks, Tianfan Xue, Nikhil Karnad, Qiurui He, Jonathan T. Barron, Dillon Sharlet, Ryan Geiss, Samuel W. Hasinoff, Yael Pritch, Marc Levoy, “Handheld mobile photography in very low light.” SIGGRAPH
Asia 2019.
• Liba, Orly, Longqi Cai, Yun-Ta Tsai, Elad Eban, Yair Movshovitz-Attias, Yael Pritch, Huizhong Chen, and Jonathan T. Barron. "Sky Optimization: Semantically Aware Image Processing of Skies in Low-Light Photography." In Proceedings of the
IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pp. 526-527. 2020.
• Chen, et al. "Learning to see in the dark." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018.
• Wronski B, Garcia-Dorado I, Ernst M, Kelly D, Krainin M, Liang CK, Levoy M, Milanfar P . “Handheld multi-frame super-resolution”, SIGGRAPH 2019.
• Ben-Ezra and Nayar, "Motion Deblurring using Hybrid Imaging”, CVPR 2003
• Yuan, Sun, Quan, Shum, “Image Deblurring with Blurred/Noisy Image Pairs”, ACM SIGGRAPH 2007
• Hasinoff, Kutulakos, “Confocal Stereo”, ECCV 2006
• Hasinoff, Kutulakos, “A Layer-Based Restoration Framework for Variable-Aperture Photography”, ICCV 2007
• Raskar, Tan, Feris, Yu, Turk, “Non-photorealistic Camera: Depth Edge Detection and Stylized Rendering using Multi-Flash Imaging”, ACM SIGGRAPH 2004
• Pettschnigg, Agrawala, Hoppe, Szeliski, Cohen, Toyama, “Digital Photography with Flash and No-Flash Image Pairs”, ACM SIGGRAPH 2004
• Eisemann, Durand, “Flash Photography Enhancement via Intrinsic Relighting”, ACM SIGGRAPH 2004