2. It is hard to know what
it cannot do!
Promise of
AI
It is hard to know
what it can do.
Performance of AI
The Gap
AI in Vision
• AI in data science applications has
been embraced by many
enterprises.
• Typical tools range from classical
statistical techniques to machine
learning, deep nets and
occasionally reinforcement
learning.
• There has been a lot of noise of AI
in IOT - where sensor data has
been treated with data science
techniques.
• However, the real challenge
comes in the Vision world. Where
data is analog, amorphous,
arbitrary and awfully heavy.
• With the advent of sight, there was
an explosion of species.
• Eyes, became a weapon in the
war of survival.
• Given vision’s importance, human
brain uses almost 35% of its
capacity to process visual inputs.
• Essentially there are two pathways
- the ventral stream that identifies
the “what” and the dorsal stream
that analyzes “where, when”
• With so much power, humans are
able to effortlessly classify, count,
recognize etc.
Vision is Messy
3. AI Vision - hybrid approach
• Given the complexity of vision
processing, it becomes important
to analyze images both at the
• Edge, and at the
• Server.
• Edge analytics is equivalent to the
ventral stream, ideal for with
scalable deployments
• Server analytics, is like the dorsal
stream - gives better correlation
and inferential remediation
• Together the Hybrid approach
solves real world problems - just
as it does in biological life.
• Intel provides an excellent balance
between Sever and Edge Centric
computing.
• While edge CPUs support inbuilt GPU
and new era Neural Accelerators, the
Servers support host of own and third
party discrete GPUS as needed.
• The high speed and high capacity
memory - for CPU and GPU further
accelerate performance.
• Finally a host of optimized frameworks -
Openvino, MediaSDK, IMKL etc speed
up mathematical and matrix operations.
Hybrid - GPU + CPU,
Edge + Server
4. Real-world AI in Vision
• Define the business problem
• Assess the vision related inputs available
• Design Capture and Illumination system
• Rolling/ global shutter
• Visible, SWIR, MWIR LWIR
• FPS, FOV
• Collect data and curate the same
• Architect and deploy Classical and Neural Net Algos for
vision processing
• Deploy, test and improve accuracy
• Optimize for performance and security.
5. Curating the data
• Data being critical, proper experiment
design to collect data becomes critical
• Mixing anomalous data with regular data is
easier said than done.
• Even creating synthetic data, needs
simulation of reality.
• Finally, based on the exact nature of the
problem data sets need to be collected.
• A sample tree structure for vision problems
is as shown.
“Data is the source code”
6. Bringing things together
• A trivial approach to AI Vision problem
solving today is - to keep adding layers to
the Net, try different loss functions and train
with greater amounts of data.
• Capture system, Algorithms and the
Compute system together deliver results in
the real world.
• We have developed multi-spectral camera
systems, worked with visible, IR, Thermal
illumination, and handcrafted algorithms to
run on Intelx86, Movidius and Openvino
platform to deliver results to clients.
7. Vision Pipeline
• Real world applications require
robust pipeline between sensor
input and algorithms
• Video frames need to be
buffered and fed to multiple
processes
• Right video frames need to be
selected for analysis
• Drumbeat across incoming
video, algos, analyzed output
needs to be maintained
• This is as important as the
algos
• Standard datasets - ImageNet,
• Architectures - ResNet, VGG,
Inception,
• CNN, RCNN, LSTM, GANs
• Frameworks - TensorFlow,
MXNet, Caffe, Keras
• Deep vs Speed, Sparse vs
Dense
AI at Work
8. Few Case Studies
• A client wanted to study audience
in a high footfall event - 200,000
people were expected in few
days.
• A client wanted to read number
on his warehousing trucks, in
very poor lighting condition.
• Another client wanted to create
night vision camera systems for
driver assistance in inclement
weather.
• EPC company wanted to monitor
progress of a project across sites
with inexpensive cameras,
instead of 3D Laser Lidars.
• Intel is an established platform
• Linux support is simply awesome
• Body of knowledge is very high - both at
the edge - embedded areas; as well as
at the server Linux OS.
• Host of optimized tools squeeze the
performance out of systems
• There are solutions for every size and
scale, unlike options in the market.
• Blend of CPU and GPU allows easy
introduction of new models, till the
implementations come on the NN
platforms.
Intel Difference
9. Outdu - Open AI Vision Platform
• Outdu has built, tested, deployed
and enterprise grade solutions
across industries
• Its multi-spectral camera systems
and host of optimized AI algorithms
in People, Vehicles and Scenes
deliver high accuracy results
• Its Robust X-OPS platform allows
Configuration, Visualization and
Orchestration of 1000s of Edge AI
devices.
• Its Open X-OPS platform allows
third party software developers and
Domain experts to deliver AI
algorithms on target devices easily.
• There are brilliant companies,
individuals and experts who create
algorithms that can solve problems.
• Enterprise ready and friendly Intel
technology becomes the fastest way
to develop AI solutions.
• Our open architecture platform allows
these algorithms to reach the edge
devices, and harness the - camera,
mic, imu, gps, and other sensors.
• Enterprises can effortlessly deploy
large scale AI solutions with Outdu.
Working Together
It is hard to know what
it cannot do!
Promise of
AI
It is possible to do.
Performance of AI
Outdu - bridging the Gap.