1
CS-ToF: High-resolution Compressive time-of-flight imaging Fengqiang - - PowerPoint PPT Presentation
CS-ToF: High-resolution Compressive time-of-flight imaging Fengqiang - - PowerPoint PPT Presentation
CS-ToF: High-resolution Compressive time-of-flight imaging Fengqiang Li, Chia-kai Yeh, Kuan He, Oliver Cossairt (Northwestern University) Huaijin Chen, Adithya Pediredla, Ashok Veeraraghavan (Rice University) fengqiang.li@u.northwestern.edu 1
2
Time of flight
Image credit: Donald Griffin
3
Time-of-flight (ToF) imaging
Goyer et al., Bulletin of the American Meteorological Society, 1963
Pulsed laser Detector Object Clock
4
Time-of-flight (ToF) imaging
Goyer et al., Bulletin of the American Meteorological Society, 1963
Pulsed laser Detector Object d Clock
5
Time-of-flight (ToF) imaging
Goyer et al., Bulletin of the American Meteorological Society, 1963
Pulsed laser Detector Object d
d = 1 2 ⋅ c ⋅ t
Clock
6
Time-of-flight (ToF) imaging
Goyer et al., Bulletin of the American Meteorological Society, 1963
Pulsed laser Detector Object d
d = 1 2 ⋅ c ⋅ t
Clock
Pulsed light based ToF
q Raster scanning
7
Continuous-wave ToF
Laser diode ToF camera Controller
!(#) %(#)
Schwarte et al., SPIE Proceeding, 1997
# = 0 !(#) r(#)
8
Continuous-wave ToF
Laser diode ToF camera Controller
!(#) %(#)
p Schwarte et al., SPIE Proceeding, 1997
# = 0 !(#) s(#) r(#) () * ( # )
()
9
10
Robotics Autonomous cars HCI VR/AR
11
Robotics Autonomous cars HCI VR/AR Non-Line-of-Sight imaging Transient imaging
12
Low spatial resolution
- 640✕480 pixels (0.3 mega) vs 120-megapixel CMOS
13
Low spatial resolution
- 640✕480 pixels (0.3 mega) vs 120-megapixel CMOS
- Extra circuits for each pixel
- Limited wafer size
14
Work directly on ToF camera output:
- Edge guidance
- Defocus debluring
Fuse ToF output with a second camera:
- RGB camera
- Stereo, Photometric stereo, Shape from polarization
Previous work
15
Optical multiplexing
Objects Spatial light modulator-SLM (High resolution) ToF camera (low resolution)
q Multiple Pixels (e.g. 3✕3) on spatial light modulator projected on One Pixel of ToF camera
16
Motivation for optical multiplexing
q Higher resolution compared to algorithm based method q No need to fusion with different image modalities q Resolution is dependent on SLM
17
Pattern on modulator (M) ToF output (y)
Compressive sensing … …
18
Phasor
Intensity (a) depth (d) q ToF output: intensity (a) and depth (d) q Phasor: a ∘ e$%&'
( ) = a ∘ e$+
19
Imaging forward model
x = !" ∘ e%&' q High resolution scene projection on modulator (x)
(
20
Imaging forward model
x = !" ∘ e%&' q High resolution scene projection on modulator (x) q Spatial light modulator pattern (M)
M
()
21
Imaging forward model
x = !" ∘ e%&' q High resolution scene projection on modulator (x) q Spatial light modulator pattern (M) q Translation matrix from modulator to ToF camera (C)
M C
()*
22
Imaging forward model
x = !" ∘ e%&' q High resolution scene projection on modulator (x) q Spatial light modulator pattern (M) q Translation matrix from modulator to ToF camera (C) q ToF camera output (y)
M
( = *+,
C
y = !- ∘ e%&.
23
Imaging forward model
!" !# !$ . . !& = (" (# ($ . . (& )
* = +,) = ()
q ToF Output y. with modulation pattern M. 01234567 01834569 01:3456; . . 01<3456= = ("0>2345?7 (#0>8345@9 ($0>:345@; . . (&0><345@A = (" (# ($ . . (& 0>345@
24
Optimization
! x = arg min
*
+ − -. / + 1Φ(x) Φ . = TV . = 7
8
G: x8
/ + G; x8 /
25
Simulation
q High resolution scene
- Middlebury 3D Datasets
- Size: 1140x912
q At – simulated system A = CMt
- Mt – Hadamard multiplexing patterns on DMD at t-th measurement
- C – mapping matrix, defined as spatial down-sampling by averaging
q Low resolution measurements
- Simulated through via yt = Atx
- Size: 120×153
26
(b). Original LR ToF measurement Intensity (a). Ground truth Intensity (c). HR reconstruction (50%) Intensity (d). HR reconstruction (25%) Intensity (e). HR reconstruction (15%) Intensity
27
(b). Original LR ToF measurement Intensity (a). Ground truth Intensity (c). HR reconstruction (50%) Intensity (d). HR reconstruction (25%) Intensity (e). HR reconstruction (15%) Intensity (g). Original LR ToF measurement Phase (f). Ground truth Phase (h). HR reconstruction (50%) Phase (i). HR reconstruction (25%) Phase (j). HR reconstruction (15%) Phase
28
Original LR ToF measurement Ground truth HR reconstruction (15%)
29
Original LR ToF measurement Ground truth HR reconstruction (15%)
30
Prototype
q Texas Instrument DLP 4500: 1140×912 pixels q ToF camera: Texas Instrument OPT 8241 320×240 (186×200 pixels)
31
Original LR ToF Measurement Pixel scanning
32
3D scene
33
Original LR ToF measurement
HR Reconstruction (no compression)
HR Reconstruction (0.6) HR Reconstruction (0.25)
Depth in m Depth in m Depth in m Depth in m
34
Native No compression CS: 0.6 CS: 0.25
35
Original LR ToF measurement HR reconstruction (no compression) HR reconstruction (60%) HR reconstruction (25%)
36
Native No compression CS: 0.6 CS: 0.25
37
Depth in mm
Ground truth depth Depth values with 25dB SNR
Depth resolution
q No depth resolution improvement q Depth resolution is better than bicubic interpolation
38
39