More Related Content Similar to "Snapdragon Hybrid Computer Vision/Deep Learning Architecture for Imaging Applications," a Presentation from Qualcomm (20) More from Edge AI and Vision Alliance (20) "Snapdragon Hybrid Computer Vision/Deep Learning Architecture for Imaging Applications," a Presentation from Qualcomm1. © 2019 Qualcomm
Qualcomm® Snapdragon™ Hybrid
Computer Vision/ Deep Learning
Architecture for Imaging Applications
Robert Lay, Product Manager, Staff
Qualcomm Technologies, Inc (QTI)
May 2019
Qualcomm Snapdragon is a product of Qualcomm Technologies, Inc. and or its subsidiaries.
2. © 2019 Qualcomm 2
Today and tomorrow’s imaging challenges
The challenges of
Camera workloads
Very compute intensive
Complex algorithms
and Data flows
Many concurrencies
Low latency,
high bandwidths
Must be thermally efficient
for sleek, ultra-light designs
Requires long battery
life for all-day use
Storage / Memory Compute/
Bandwidth limitations
3. © 2019 Qualcomm
Extreme
Speed Boost
Up to 4x*
Power Savings
Multi-object Classification
Multi-Object Tracking
Object Segmentation
Depth Sensing at 60fps
6DoF XR Body Tracking
CV Stabilization
QTI has enabled the world’s first announced CV-ISP
Computer Vision HW Acceleration
A Powerful Heterogeneous Platform
for Advanced Imaging Experiences
*vs running on cDSP, Q6
Qualcomm Adreno, Qualcomm Hexagon, Qualcomm Kryo, Qualcomm Spectra and Qualcomm Snapdragon are products of Qualcomm Technologies Inc., and/or its
subsidiaries.
4. © 2019 Qualcomm 4
Heterogeneous Approach at QTI to Computer Vision
ISP
Algorithmic
advancements
Neural network algorithm design
optimized for hardware
Optimization for space and
runtime efficiency
Software
tools
PC and target based simulators
Flexible SW frameworks
Sample code
Efficient
hardware
Efficient architecture design
Selecting the right compute
engine for the right task
5. © 2019 Qualcomm
Qualcomm® SpectraTM CV HW Processor - CVP
CVP Acceleration Applications and Benefits
Preprocessing
& Detection
• Pyramidal scaler
• Downscaler
• Feature descriptor
Tracking/
Motion processing
• Optical flow
• Feature point extraction
• Binary feature descriptor and inline descriptor matching
• Normalized cross correlation (NCC)
Classification • Non-linear classifier
Geometric Correction
Engine
• Low-power high-quality image warping and alignment engine
Depth Processing • Depth from Stereo
Embedded Processor • Scheduling and concurrency handling
7. © 2019 Qualcomm
Example for CVP accelerated real time video
bokeh/segmentation
Blue blocks accelerated using CVP
R+L
Images
RT
Bokeh
DL Post-Proc
Depth from
Stereo
DFS
Rectification
GCE, Scalers
DL Post Proc
Blur Map
DL
Segmentation
RT
Segmented
RT
Recalibration
Descriptors, OF
Scalers
Factory
Calibration
8. © 2019 Qualcomm
Depth sensing at 60fps
World’s First announced 4K HDR
with portrait mode
Example for CVP accelerated
real time video bokeh
9. © 2019 Qualcomm
XR Case:
CVP acceleration for 3D scene understanding
Book:
Ireland, ISBN,
<url@amazon>
Table: IKEA #,
dimensions W x H x D
unknown:
3D model
Unknown:
cube W x H x D
Known 3D coordinates
of everything
Room floor plan
and furniture layout
10. © 2019 Qualcomm
Blue blocks accelerated using CVP
3D reconstruction
Camera pose
Geometry
Semantic
information
…
Viewer perspective
6DOF
Depth
Image
Data
Planes
Per object
extraction
Room
layout
3D
Segmentation
DL
fusion
DL semantic
segmentation Obstacle
warningIR
RGB
Mono
11. © 2019 Qualcomm 11
XR 6DoF accelerated with CVP
Feature Detector and Tracker
(Major tasks offloaded to CVP)
Extended Kalman Filter
(Done in cDSP and optionally mDSP)
Predicted
2D feature
positions
Measured
2D feature
position
Pose
New
Feature
Inertial
Sensor
calibration
Pose
3D
Map Gravity
Camera
Frames
Camera intrinsic &
extrinsic calibration
Accelerometers
Gyroscopes
12. © 2019 Qualcomm
XR 6DoF accelerated with CVP
CVP
Left
Camera
Frames
Right
Camera
Frames
DSP
Pyramid Scaler
3 Levels
Feature
Tracking
3 Levels NCC
Keypoint
Detection
Predict
features
location
Camera
Pose
EKF
Update
Key
Frame?
Stereo
Matching
NCC
Coordinates Coordinates
13. © 2019 Qualcomm
6DoF Body Tracking Object Detection and Tracking
CPU+DSP+GPU CVP
2xPower
Savings
4xPower
Savings
Power savings with CVP*
*Qualcomm R&D characterization on Snapdragon 855 mobile platform
14. © 2019 Qualcomm
6DoF Body Tracking Object Detection and Tracking
CPU+DSP+GPU CVP
39%Compute
Savings
96%Compute
Savings
Compute savings with CVP *
*Qualcomm R&D characterization on Snapdragon 855 mobile platform
15. © 2019 Qualcomm 15
Conclusions
Computer Vision
underpins cutting edge
imaging experiences
Classic camera and video
experiences enhanced with:
• Motion compensated temporal filtering
• Multi-frame techniques
• Depth enabled features and algorithms
• Video superresolution
• Video frame rate upconversion
Emerging industries enabled by vision:
QTI is taking a heterogeneous
approach to computer vision enhanced
camera experiences
Meeting KPIs by combining general purpose
compute engines with dedicated accelerators:
IVI/ADASXR IoT
Power Latency Processor
and memory
capacity
16. © 2019 Qualcomm 16
Resources
Snapdragon Computer Vision SDK
Please contact Robert Lay: rlay@qti.qualcomm.com
FastCV Computer Vision SDK
https://developer.qualcomm.com/software/fastcv-sdk
Snapdragon On-Device AI
https://www.qualcomm.com/invention/artificial-intelligence
https://www.qualcomm.com/snapdragon/artificial-intelligence
Qualcomm® Neural Processing SDK
https://developer.qualcomm.com/software/qualcomm-neural-processing-sdk
Hexagon SDK
https://developer.qualcomm.com/software/hexagon-dsp-sdk
Qualcomm Neural Processing SDK is a product of Qualcomm Technologies, Inc. and/or its subsidiaries.
17. © 2019 Qualcomm
CVP SDK for Customer Enablement
• Documentation:
• CVP 1.0 (SD855) API Document: Available now
• Computer Vision Processor (CVP) v 1.x Programming Guide
• CVP 2.0 (SD865) API Document: May 2019
• Computer Vision Processor (CVP) v 2.x Programming Guide
• Bit-exact PC simulator (Windows and Linux) for algorithm development:
• CVP 1.0 simulator: Available now
• CVP 2.0 simulator: Q2/2019)
• Reference examples code for each of CVP blocks: CVP1.0 now, CVP2.0 Q2/2019
• Offline training tool for Object Detection: CVP1.0 now, CVP2.0 Q3/2019
• This tool works with annotations same as used for training DL networks
18. © 2019 Qualcomm
Thank you
Nothing in these materials is an offer to sell any of the components or devices referenced herein.
Qualcomm, Snapdragon, Adreno, Hexagon, Kryo and Qualcomm Spectra are trademarks of Qualcomm
Incorporated, registered in the United States and other countries. Other products and brand names may be a
trademark or registered trademark of their respective owners.
References in this presentation to “Qualcomm” may mean Qualcomm Incorporated, Qualcomm Technologies,
Inc., and/or other subsidiaries or business units within the Qualcomm corporate structure, as applicable.
Qualcomm Incorporated includes Qualcomm’s licensing business, QTL, and the vast majority of its patent
portfolio. Qualcomm Technologies, Inc., a wholly-owned subsidiary of Qualcomm Incorporated, operates, along
with its subsidiaries, substantially all of Qualcomm’s engineering, research and development functions, and
substantially all of its product and services businesses, including its semiconductor business, QCT.