This new release brings Image Processing to the BigML platform, a feature that enhances our offering to solve image data-driven business problems with remarkable ease of use. Because BigML treats images as any other data type, this unique implementation allows you to easily use image data alongside text, categorical, numeric, date-time, and items data types as input to create any Machine Learning model available in our platform, both supervised and unsupervised.
Now, it is easier than ever to solve a wide variety of computer vision and image classification use cases in a single platform: label your image data, train and evaluate your models, make predictions, and automate your end-to-end Machine Learning workflows. As with any other BigML feature, Image Processing is available from the BigML Dashboard, API, and WhizzML, and it can be applied to solve use cases such as medical image analysis, visual product search, security surveillance, and vehicle damage detection, among others.
2. BigML Release: Image Processing
BigML, Inc 2
Image Processing Release Webinar
Enter questions into chat box. We will answer some via chat and others
at the end of the session.
QUESTIONS
info@bigml.com @bigmlcom
Atakan Cetinsoy
VP of Predictive Applications.
MODERATOR
Charles Parker Ph.D.
VP of Machine Learning Algorithms.
SPEAKER
https://bigml.com/releases/image-processing
RESOURCES
3. BigML Release: Image Processing
BigML, Inc 3
BigML Image Processing
• ALL the tasks for generating insights from image data
ON A SINGLE PLATFORM: From labeling to inference,
evaluation, and predictions.
• Streamlined image dataset management with
composite sources.
• Comprehensive feature extraction options to feed any
algorithm.
• Pre-trained CNNs for classification and regression.
• Automatic handling of infrastructure concerns.
• Your choice: Code and no code.
Solving image data-driven business problems with remarkable ease of use
5. BigML Release: Image Processing
BigML, Inc
In This Webinar
Some examples of full applications using image processing
1
2
3
4
A simple image classifier in BigML
Image data management on the platform
Some non-trivial examples of image processing problems
5
6. BigML Release: Image Processing
BigML, Inc
(A Few) Applications of Computer Vision
• Ecological: Land use classification,
ecosystem health management, animal
identification.
• Transportation: Tollway monitoring,
anomalous event detection, automated
billing.
• Medical: Medical image classification,
segmentation for disease identification.
• Finance: Automated check reading, form
classification, security, insurance claim
estimation.
6
7. BigML Release: Image Processing
BigML, Inc 7
Application #1 - License Plate Monitoring
• Globalvia tollway monitoring:
• Object detection to pull plate
image from car.
• Object detection for pulling
characters / symbols from plate.
• Classifier for deciding what state
the plate is from.
• The combination of all of the
models with bespoke business
rules lead to a successful system.
8. BigML Release: Image Processing
BigML, Inc 8
Application #2 - Marksmanship Training
• Accushoot Marksmanship Training Application:
• Object detection: Target detection, gun detection.
• Image Classification: Grip classification, gun
classification, target classification, impact
classification.
• Other models: Classifier for shooting phases
(drawing, shooting, holstering, reloading).
• Again, making a great application with
computer vision often takes more than one
model.
14. BigML Release: Image Processing
BigML, Inc 14
How is BigML Different?
• Most competitors that provide learning over images focus on a mapping of
image -> class, but there’s so much more!
• On the input side:
• We can extract meaningful features rather than using the raw pixel data.
• We can combine the image with other features (captions, geo-coordinates, etc.).
• On the output side:
• We can do regression.
• We can do unsupervised learning.
Image processing on BigML gives the user the flexibility, traceability,
and scalability that they get with any other datatype.
15. BigML Release: Image Processing
BigML, Inc
So Deep Learning, right?
• That’s an important part of our offering,
but not the only part.
• Yes, if you do select a BigML deepnet and
have image data, you’ll get a CNN.
• We try several common architectures and
select the best one.
• Deep learning is great, and is very often
the best thing available, but sometimes
it isn’t.
15
16. BigML Release: Image Processing
BigML, Inc 16
Main Advantages of BigML
• User labeling and image dataset
management with composite sources.
• A variety of image feature extraction
techniques that can be used to feed any
BigML algorithm.
• Full training of convolutional neural
networks for classification and regression.
• Handling of infrastructure concerns for
image training (which is not trivial or cheap).
17. BigML Release: Image Processing
BigML, Inc 17
Composing Image Datasets
• Zip files are nice, but what if you need to
compose an image dataset, and add to it
incrementally?
• For this we offer image labeling and
composite sources:
• A composite source is a source that contains
a bunch of sub-sources (individual images).
• While the composite source is open, you can
add images, or labels to images, which
preserves immutability.
19. BigML Release: Image Processing
BigML, Inc 19
Adding Metadata
• What if there was other information?
• GeoCoordinates where the photo was
taken.
• Approximate altitude.
• Adding that information with BigML is as
simple as editing the CSV.
• The image is just some of the features
in the instance.
21. BigML Release: Image Processing
BigML, Inc
$
Car image
Features
Regression model
21
Applications: Insurance Claim Estimate
• Suppose you’re an auto insurance company.
• Develop software where you have customer take a picture of the damage to
estimate the severity of the claim.
• Simple regression problem, image -> $
22. BigML Release: Image Processing
BigML, Inc
$
Front image
Features
Side image
Features
Brake pedal
State Steering wheel
Angle
22
• Now, what if you have them take both a front and side image?
• What if you have telemetry data from the car?
Applications: Insurance Claim Estimate
23. BigML Release: Image Processing
BigML, Inc
$
Word counts Accelerometer
Screen on?
Front image
Features
Side image
Features
Brake pedal
State Steering wheel
Angle
23
• How about the text of the police report?
• User’s mobile phone activity in the
moments before the crash?
Applications: Insurance Claim Estimate
24. BigML Release: Image Processing
BigML, Inc 24
Unsupervised Learning
• How do you do anomaly detection or clustering with
images?
• Change each image to a (somewhat) short vector of numbers.
• Treat each of these vectors as a set of numeric features for
each instance.
• These features can be combined with any other feature type.
• Deepnets do this implicitly, but other ML algorithms require
an explicit choice.
• By default, we choose something that “often works”, but
this is an opportunity to encode domain knowledge.
26. BigML Release: Image Processing
BigML, Inc
Simple Model
More Complex
Most Complex
Reject
Pass
Pass
Accept
26
Model Cascade
• Those features can feed supervised models,
too, and may be much faster.
• Use an increasingly accurate and increasingly
expensive chain of models as “gates”:
• Use a model that’s fast, has near-perfect recall,
but poor precision.
• If something passes the first model, pass it on to
a fancier one with better precision.
• Repeat as necessary.
• In this way, only a few instances require the
most time consuming prediction.
Reject
Reject
27. BigML Release: Image Processing
BigML, Inc 27
Image Processing the BigML Way
• Everything is immutable, traceable, and composable
• All of your models are always available and never change
• You can always see exactly the training data that generated
the model
• Everything is available via the API and downloadable
• Image dataset creation and labeling can be done via the API
• Models are live via API endpoint as soon as they are created
• Models can be downloaded and run locally
28. BigML Release: Image Processing
BigML, Inc 28
Okay, Images are A Little Bit Special
• I’ve glossed over the fact that there are some
problems that are specific to images:
• Object detection
• Image segmentation
• Superresolution
• Our initial release won’t support these functionalities
immediately, but object detection is coming.
• Accushoot, Globalvia, and several other customers
are using the beta version of object detection right
now.