Upload
jiwei-li
View
31
Download
1
Embed Size (px)
Citation preview
Make Smartphones Last A Day: Pre-processing
Based Computer Vision Application Offloading
Jiwei Li*, Zhe Peng*, Bin Xiao*, Yu Hua**
The Hong Kong Polytechnic University*Huazhong University of Science and Technology**
Outline• Background & Motivation• Pre-processing Based Offloading• Evaluation• Conclusion
Background & Motivation
Computer Vision Applications
Object Recognition
3D Reconstruction
Augmented Reality
Power hungry Compute intensive
Offloading Techniques• To tackle limitations of smartphones• Limited computational power• Limited battery life
• Offloading compute-intensive jobs to remote servers via wireless networks• Improve response time• Improve energy efficiency
Offloading Image Processing
Original Image (8 Mb)
Processed Image(8+ Mb)
Transferring images consumes a lot of energy.
Original Image (8+ Mb)
Original Image (2+ Mb)
Pre-processing helps
Processed Image(2+ Mb)
Why pre-processing?• Less active time on the network interface• Save energy
• Drawbacks• Compromise the image processing accuracy• Pre-processing incurs energy cost
Challenges• Need to know how exactly pre-processing will
affect the energy consumption and final processing accuracy• Important for determining the right pre-processing level
• Need to work at both static and dynamical environments. • Predication-based techniques should be avoided
Motivating Experiments• Understand the effect of pre-processing
• Face detection experiments:• We adopt downscaling as the pre-processing technique• Correct detection rate, false positive rate• File size
Effect of Downscaling
As the downscaling factor decreases from 1 to 0.1:• the average correct detection rate drops• the average false positive rate drops• the file size drops linearly
Smaller image file sizes indicate lower result accuracy.
The Studied Problem• Energy consumption as a constraint• E.g., one day
• Overall response time as a constraint
• Problem: maximize the final processing accuracy subject to energy and time constraints
People charge their phones every day
Pre-processing Based Offloading
Overview
Two decisions:• When to start data transmission via which network interface• Determine the pre-processing level
Decisions are made based on current channel information and accumulated delay time.
Single Task Offloading• Calculate the condition check deadline• Determine when to perform condition check
periodically (two parameters: D(0) and ) • Determine the image pre-processing level
Maximize Final Processing Accuracy
• (1) is time constraint that the task must be finished before completion deadline • (2) is power constraint that the task consumes no more energy than using cellular
Image file size
Minimum image file size
Remaining time till conditioncheck deadline
Minimize pre-processing level
Analytical Framework• Study how Smax is affected by and D(0)• Assume Bw follows an exponential distribution with
rate λ• Calculate expected Smax
Expected Smax
Multiple Task Offloading• Task execution order: FIFO• Recalculate the new condition check deadline
Evaluation
Implementation• Client side (Android smartphone):• Estimate the TCP throughput of current connected Wi-Fi• Downscale images
• Server side (Windows Server 2012 R2):• Process images using OpenCV library
Experiment Methodology• Face detection
• Static environment• Compare different pre-processing levels• Compare Wi-Fi and 3G environments
• Dynamic environment• Compare different D(0) and β• Compare against other approaches
Static Environment – Different Wi-Fi Environments
• The energy consumption decreases when the downscaling factor decreases.
• Some exceptions• Better Wi-Fi results in lower energy consumption.
Static Environment –Comparison of Wi-Fi & 3G
Dynamic Environment –Image Processing Accuracy
Dynamic Environment –Comparison with Other Approaches
• Compare against• Instant upload • Instant check
Conclusion• Understand how the image pre-processing level would
affect accuracy and energy consumption on smartphones
• Propose a Pre-Processing Based Computer Vision Application Offloading approach
• Achieves near-optimal image processing accuracy while satisfying energy and time constraints
Q&A
Thank you