Upload
trinhmien
View
218
Download
4
Embed Size (px)
Citation preview
June 2, 2017
S.LSI ICRA'17 Workshop on "Event-based Vision"
The Road to Market
Dynamic Vision Sensor
Yoel Yaffe*, Nathan Levy, Evgeny Soloveichik, Sebastien Derhy, Ayal Keisar, Elad Rozin, Liron Artsi
Jun-Seok Kim†, Keunju Park, Bongki Son, Yunjae Suh, Heejae Jung, Changwoo Shin, Jooyeon Woo, Yohan Roh, Hyunku Lee, Hyunsurk (Eric) Ryu
*Samsung Israel Research Center, Israel †System LSI , South Korea
Samsung Electronics
Contact: [email protected]
2 S.LSI ICRA'17 Workshop on "Event-based Vision"
The Road Less Traveled
image by aosleading
3 S.LSI ICRA'17 Workshop on "Event-based Vision"
Market Driven Requirements VS Early DVS Prototypes
Requirements Early DVS Prototypes
Low Cost Large Pixel No standard for Quality / Die sorting
Minimal Module Size Large, due to large Pixel -> Large optical format
Ultra Low Power Low power
Good Event Quality Redundant events (especially at low light) Motion artifacts, Timing accuracy Noise, Flicker
Low Data Rate Scene dependent; In some cases may exceed CIS typical throughput
Vision processing at Edge Not available
4 S.LSI ICRA'17 Workshop on "Event-based Vision"
The Road to Market - Overview
oReducing Cost and Module Size
oReducing Power
oDelivering Good Event Quality
oData Throughput Reduction
oEvent Processing Acceleration
oSummary
5 S.LSI ICRA'17 Workshop on "Event-based Vision"
The Road to Market - Overview
oReducing Cost and Module Size
oReducing Power
oDelivering Good Event Quality
oData Throughput Reduction
oEvent Processing Acceleration
oSummary
6 S.LSI ICRA'17 Workshop on "Event-based Vision"
Reducing Pixel Size and Optical format
18.5 x 18.5 µm2 → 6 x 6 µm2 ~90% pixel area reduction
Pixel Shrink
o BSI process
o Downsizing MIMCAP Without a decrease in minimum contrast sensitivity
o Stacked Cu-Cu Technology
6 µm
9 µm
12 µm
18.5 µm
by Stacked Pixel
Technology
PD + LOGA+C1
C2+CFA+Comp.
7 S.LSI ICRA'17 Workshop on "Event-based Vision"
Reducing Pixel Size and Optical format
3μm x 3μm Photodiode / 9μm x 9μm Pixel
5μm x 5μm Photodiode / 6 μm x 6 μm Pixel / Wafer-stacking
o Photo Enhancement:
8 S.LSI ICRA'17 Workshop on "Event-based Vision"
Delivering Product to Market - Overview
oReducing Cost and Module Size
oReducing Power
oDelivering Good Event Quality
oData Throughput Reduction
oEvent Processing Acceleration
oSummary
9 S.LSI ICRA'17 Workshop on "Event-based Vision"
Reducing Power
oUse Samsung’s advanced process
o Subsampling modes
oDynamic wakeup, window of interest
10 S.LSI ICRA'17 Workshop on "Event-based Vision"
Reducing Power: Statistics for Dynamic Decisions
o Spatial Histogram & Blocking Area:
ISSCC’17
32
32
(0,0) (0,1) (0,2) (0,3) (0,16) (0,17) (0,18) (0,19)
(1,0) (1,1) (1,2) (1,3) (1,16) (1,17) (1,18) (1,19)
(14,0) (14,1) (14,2) (14,3) (14,16) (14,17) (14,18) (14,19)
640x480 Image Plane
11 S.LSI ICRA'17 Workshop on "Event-based Vision"
Reducing Power: Statistics for Dynamic Decisions
ISSCC’17
Spatial Histogram Data
12 S.LSI ICRA'17 Workshop on "Event-based Vision"
Reducing Power: Blocking, Subsampling
ISSCC’17
SNE
AER
APS+ANALOG
APS
AN
ALO
G
Cropper
AER
AER Packetizer
AER
FIFO
AER Packet
MIPI I/F
USB I/F
Parallel I/F
AER
Pack
et
OUTPUT
MUX
Spatial Histogram
IIC I/F
PAD Out
AER
Histogram
INT I/FPAD Out
IIC Bus
Wakeup
Histogram
Interrupt
Histogram
Digital AER
AER Gen
Analog Test Control
Time Stamper
Test Pattern
Subsampler
Blocking Area
13 S.LSI ICRA'17 Workshop on "Event-based Vision"
Reducing Power: Statistics Collection
ISSCC’17
SNE
AER
APS+ANALOG
APS
AN
ALO
G
Cropper
AER
AER Packetizer
AER
FIFO
AER Packet
MIPI I/F
USB I/F
Parallel I/F
AER
Pack
et
OUTPUT
MUX
Spatial Histogram
IIC I/F
PAD Out
AER
Histogram
INT I/FPAD Out
IIC Bus
Wakeup
Histogram
Interrupt
Histogram
Digital AER
AER Gen
Analog Test Control
Time Stamper
Test Pattern
Subsampler
Blocking Area
14 S.LSI ICRA'17 Workshop on "Event-based Vision"
Reducing Power: Wakeup
ISSCC’17
SNE
AER
APS+ANALOG
APS
AN
ALO
G
Cropper
AER
AER Packetizer
AER
FIFO
AER Packet
MIPI I/F
USB I/F
Parallel I/F
AER
Pack
et
OUTPUT
MUX
Spatial Histogram
IIC I/F
PAD Out
AER
Histogram
INT I/FPAD Out
IIC Bus
Wakeup
Histogram
Interrupt
Histogram
Digital AER
AER Gen
Analog Test Control
Time Stamper
Test Pattern
Subsampler
Blocking Area
15 S.LSI ICRA'17 Workshop on "Event-based Vision"
The Road to Market - Overview
oReducing Cost and Module Size
oReducing Power
oDelivering Good Event Quality
oData Throughput Reduction
oEvent Processing Acceleration
oSummary
16 S.LSI ICRA'17 Workshop on "Event-based Vision"
Good Event Quality: Objective
Clean and perfect edge map Obtained by subtracting frames on a High Frame-Rate CIS camera
17 S.LSI ICRA'17 Workshop on "Event-based Vision"
Good Event Quality
Generated Events
“Meaningful” Events
18 S.LSI ICRA'17 Workshop on "Event-based Vision"
Generated Events
“Meaningful” Events
Good Event Quality
Redundant Events o Falsely Generated o Multiple detections o Not applicable
Examples: o Noise; Bad Pixels o Response to high
contrast change o Light source flicker;
Background motion ; Too many events
Missing Events o Not/badly detected o Discarded
Examples: o Overflow,
Timestamp errors o Filtered (mitigating
redundant events)
19 S.LSI ICRA'17 Workshop on "Event-based Vision"
Generated Events
Good Event Quality
Redundant Events o Falsely Generated o Multiple detections o Not applicable
Examples: o Noise; Bad Pixels o Response to high
contrast change o Light source flicker;
Background motion ; Too many events
“Meaningful” Events
Missing Events o Not/badly detected o Discarded
Examples: o Overflow,
Timestamp errors o Filtered (mitigating
redundant events)
20 S.LSI ICRA'17 Workshop on "Event-based Vision"
Redundant Events: Bad Pixels; Noise
Basic model:
o Bad Pixel: Specific Pixels that generate events at high frequency (regardless of the scene)
o Noise: Random pixels that generate events at low frequency (regardless of the scene)
Actual behavior:
Arbitrary separation between bad pixels and noise 10% of max number of events
Number of pixels per event rate
Events/sec
Event Histogram: DVS filming a white wall
21 S.LSI ICRA'17 Workshop on "Event-based Vision"
Redundant Events: Bad Pixels; Noise
Bad Pixel Suppression Example:
22 S.LSI ICRA'17 Workshop on "Event-based Vision"
Redundant Events: Bad Pixels; Noise
Noise Suppression Example:
23 S.LSI ICRA'17 Workshop on "Event-based Vision"
Redundant Events: Multiple detections
LogAMP of DVS Gen2 (ISSCC’17)
MN2
MLOG2 MSFVLOG
VSF
IPD
ILOG
MLOG1
MN1
ISF
IPD
time
VLOG
time
ON
time
nRST
time
ΔVTH_ON
ΔVTH_ON
step response
o Overall pixel BW (speed) is limited by LogAMP BW
o LogAMP BW is mainly affected by 1) Its structure; 2) The amount of photocurrent
o Redundant events generated at low light (IPD) on @ large contrast change
24 S.LSI ICRA'17 Workshop on "Event-based Vision"
Redundant Events: Multiple detections
IPD
time
VLOG
time
ON
time
nRST
time
ΔVTH_ON
ΔVTH_ON
step response
o Overall pixel BW (speed) is limited by LogAMP BW
o LogAMP BW is mainly affected by 1) Its structure; 2) The amount of photocurrent
o Redundant events generated at low light (IPD) on @ large contrast change
Solutions:
o Structure: Increase ~x5-x10 LogAMP BW by new structure
o Limit photocurrent: Introduce “HOLD” signal to pixel
“Hold” function
25 S.LSI ICRA'17 Workshop on "Event-based Vision"
Redundant Events: Light Source Flicker
o Light Source Flicker: Events that are not generated by motion but by flickering lights (with different frequencies)
o Note: This has to be differentiated from periodic motion (for example a waving hand)
26 S.LSI ICRA'17 Workshop on "Event-based Vision"
Redundant Events: Light Source Flicker
Flicker Suppression Example:
27 S.LSI ICRA'17 Workshop on "Event-based Vision"
Missing Events: Overflow, Timestamp Errors
o Overflow: Number of events describing the scene change, exceeds the circuit maximal throughput.
25% active area, 1ms refresh rate:
90% active area, 1ms refresh rate:
#1040 #1041 #1042
#109 #110 #110
28 S.LSI ICRA'17 Workshop on "Event-based Vision"
Missing Events: Overflow, Timestamp Errors
o Timestamp Errors: Events grouped according to incorrect timestamp
0
200
400
600
609556500 609557000 609557500 609558000 609558500 609559000 609559500
Frame_9 Frame_10column
time
29 S.LSI ICRA'17 Workshop on "Event-based Vision"
Solution: Pixel Readout Revision
o Solution: Number
640
… Pixel Array
480
… Prev.: Unfair arbitration, no frame context
Column selection
Column Arbiter
Readout controls Gen1 & Gen2 Gen3
Column selection Unfair arbitration scheme Scanning scheme
In-pixel memory Free running (no event hold) Free running (no event memory)
Global hold (all pixels in pixel array)
Pixel reset
Pixels with an event in a column Pixels with an event in a column
- All pixels in a column
- Global reset (all pixels in pixel array)
New: Scanning scheme, frame context options
640
… Pixel Array
480
Column selection
Scanning order : 0 to 639
30 S.LSI ICRA'17 Workshop on "Event-based Vision"
Pixel Readout: Timestamp Error Comparison
Average Event rate
Tim
est
amp
Err
or
(STD
)
Input: Poisson random events
100M 1G
1 msec
10 usec
100 usec
10M 1M
1 usec
Selective scan without frame buffer
Selective scan with frame buffer
Full scan without frame buffer
Full scan with frame buffer
Full scan with In-Pixel memory without frame buffer
31 S.LSI ICRA'17 Workshop on "Event-based Vision"
The Road to Market - Overview
oReducing Cost and Module Size
oReducing Power
oDelivering Good Event Quality
oData Throughput Reduction
oEvent Processing Acceleration
oSummary
32 S.LSI ICRA'17 Workshop on "Event-based Vision"
Data Throughput: DVS Low Event Rate Example
Less Events Easier to process
0
5
10
15
20
25
30
35
CIS 100fps CIS 30fps Moving DVS Static DVS + Strongmotion
Static DVS + weakmotion
Static DVS, no motion
Mill
ion
s
Event rate per second
Eps
33 S.LSI ICRA'17 Workshop on "Event-based Vision"
Data Bandwidth @ 32 bit/event Encoding
DVS AER 32bit per event
0
50
100
150
200
250
300
350
CIS 100fps CIS 30fps Moving DVS Static DVS + Strongmotion
Static DVS + weakmotion
Static DVS, no motion
Mill
ion
s
Bits per seconds
Bps
34 S.LSI ICRA'17 Workshop on "Event-based Vision"
Group Address Event Representation (GAER)
X
Y event
640
480
1
2
3
4
59
60
…
0
1
2
3
4
5
6
7
group_event
(0,0)
… …
AER (32 bits/event)
Address Event Representation (TS, X, Y, event) o P: packet encoding (2bit) o TS : timestamp (10bit) o X : column address : 0 ~ 639 (10bit) o Y : row address : 0 ~ 479 (9bit) o event : 1-bit. on/off
GAER (~64-4 bits/event – content dependent) Group AER (TS, X, G, group_event)
o P: packet encoding (10bit group, 12bit per col) o TS: timestamp (10bit) o X : column address : 1 ~ 640 (10bit) o G : group_address : 1 ~ 60 (6bit) o group_event : a bundle of 8 events. 16-bits
o on_evnet : 8-bits o off_event : 8-bits
35 S.LSI ICRA'17 Workshop on "Event-based Vision"
Data throughput: AER 32 bit/event
DVS AER 32bit per event
0
50
100
150
200
250
300
350
CIS 100fps CIS 30fps Moving DVS, 1000 fps Static DVS + Strongmotion
Static DVS + weakmotion
Static DVS, no motion
Mill
ion
s
Bits per seconds
Bps
36 S.LSI ICRA'17 Workshop on "Event-based Vision"
Data throughput: GAER ~12 bit / event
DVS GAER 12bit per event (average)
0
50
100
150
200
250
300
350
CIS 100fps CIS 30fps Moving DVS, 1000 fps Static DVS + Strongmotion
Static DVS + weakmotion
Static DVS, no motion
Mill
ion
s
Bits per seconds
Bps
37 S.LSI ICRA'17 Workshop on "Event-based Vision"
Data throughput: Improved GAER ~8 bit / event
DVS GAER with offset based column 8 bit per event (average)
0
50
100
150
200
250
300
350
CIS 100fps CIS 30fps Moving DVS, 1000 fps Static DVS + Strongmotion
Static DVS + weakmotion
Static DVS, no motion
Mill
ion
s
Bits per seconds
Bps
38 S.LSI ICRA'17 Workshop on "Event-based Vision"
The Road to Market - Overview
oReducing Cost and Module Size
oReducing Power
oDelivering Good Event Quality
oData Throughput Reduction
oEvent Processing Acceleration
oSummary
39 S.LSI ICRA'17 Workshop on "Event-based Vision"
Event Processing Acceleration
o GPU/DSP are not optimized for event-based computer vision:
o Considering HW accelerator/processor with a DVS-specific architecture
GPU/DSP Requirements for Event Processing
Optimized for High arithmetic density and high throughput
Low latency
Memory access Privilege coalesced data Optimized for random access
40 S.LSI ICRA'17 Workshop on "Event-based Vision"
Concept: Less but More Informative Events
o Programmable feature
o A feature event is sent by the sensor when the feature appears strongly
o For example: direction detection (can be used to produce Histogram of Gradients in the Host for human detection)
Horizontal (yellow) and vertical (cyan): Slash (red) and Backslash (green):
Applied on Human captured is a DVS movie:
41 S.LSI ICRA'17 Workshop on "Event-based Vision"
The Road to Market - Overview
oReducing Cost and Module Size
oReducing Power
oDelivering Good Event Quality
oData Throughput Reduction
oEvent Processing Acceleration
oSummary
42 S.LSI ICRA'17 Workshop on "Event-based Vision"
Summary: Market Driven Requirements
Requirements Samsung Early Prototypes Samsung DVS Product Roadmap (2017,2018)
Low Cost Large Pixel No standard for Quality / Die sorting
Similar size to 16-8MP mobile camera
Minimal Module Size Large, due to large Pixel -> Large optical format
Smaller, Optical size reduced to 1/3” (VGA)
Ultra Low Power Low power Lower power
Minimal Optical Format Large, due to large Pixel Reduced to 1/3” (VGA)
Good Event Quality
Redundant events (especially at low light) Motion artifacts, Timing accuracy Noise, Flicker
Improved pixel circuit. On chip Bad Pixel, Noise and flicker suppression. External Triggered Scan, Global reset & hold capable
Low Data Rate Scene dependent Might exceed CIS typical throughput
Compressed format in the bounds of slow CIS typical throughput
Edge device vision processing Not available Considering low level features
43 S.LSI ICRA'17 Workshop on "Event-based Vision"
Thank You!