You are on page 1of 26

Computer

Vision Solutions
About Red Buffer

Red Buffer is a premier Machine Learning and Red Buffer is uniquely positioned to provide end
Artificial Intelligence software services company. to end solutions from the development of the
machine learning engine, full stack, web and
It has worked with companies in the financial mobile applications in house.
services, legal tech, energy and other verticals
to design products and solutions specializing in These products and solutions developed by Red
predictive analytics, natural language Buffer are used by top companies including
processing and computer vision. Intuit, Dropbox, HomeAway (Expedia), Kroger,
Diageo and Under Armour.
The projects include cost and usage forecasting,
document processing and review, financial
predictions and analysis of satellite imagery.

Commercial in Confidence
Application Areas
Retail Automotive Healthcare Industrial

Visitor Heatmaps, Area of Security & Authentication Symptomatic Detection Component Inspection
Interest targeting Solutions Systems Systems

Activity recognition, auto Operating Room Manufacturing Line


check-outs Assistive Driving Features automation production count &
anomaly detection
Product Information Preliminary Diagnosis
Management automation Smart Control Systems speed-up

Commercial in Confidence
Application Areas
Agriculture Quality Assurance People Operations Sports

Crop plantation pattern Autonomous workflows Targeted Recruitment Automated Human bio-
recognition automation mechanics analysis
Robotic Process
NDVI based analytics and Automation Facial Recognition Events analytics and
crop change detection Attendance systems predictions engines
Defects / Anomalies auto-
Precision Agriculture, detection across devices Cognitive Decision Fan digital engagement
optimal resource planning and platforms Support Systems avenues

Commercial in Confidence
We have worked with several interesting
Computer Vision cases
Here are a few...
AI Assisted Golf Trainer
Playing good golf is almost a technical affair. The right
posture, swing angle and club flow among a host of
parameters differentiate a professional player from a
novice. Personal trainers come with high costs and
availability issues.

Red Buffer’s AI Assisted Golf trainer uses Computer


Vision to detect and compare player postures versus
recommended ones (benchmarked against
professionals) and provides recommendations for game
improvement.

The system does this by capturing and analyzing fine


details like joint positions, club points and wrist angles to
provide specific advice. The framework can also be used for a variety of
domains requiring human pose estimation e.g.
Toolkits used: PoseNet, current package available as fitness, fashion retail, security and CGI / AR
Android App. applications
Commercial in Confidence
Cell Clustering Detection

Cell clusters (or pieces of tissue) are


important for ensuring that there is
enough tissue present for determining
the presence of cancer cells.

We used an object detection technique


to find the cluster formation in the slide
images. Specifically, we used Retina loss
with ResNet architecture to find these
clusters.

Toolkit used: Python, Keras, Keras-


Retinanet, OpenCV

Commercial in Confidence
Cell Counting
Problem

In many areas of biomedical research including clinical pathology, cell counts obtained from images are
crucial data for diagnosing patients or for addressing hypotheses about developmental or pathological
processes. Manual cell counting is challenging in that it typically requires a specialist such as a biologist
or a pathologist to identify and characterize different cell types. Even then, manual cell counts are
subjective.

Solution

We used a deep pre-trained convolutional neural network to train the cell count problem. It was a
regression problem and the input was an image of slide and the output was the number of cells in the
slide

Toolkit used: Python, Keras, OpenCV

Commercial in Confidence
Social Distance Measurement

For a Covid-19 related use case, we have


developed a system to estimate the
effectiveness of social distancing in spaces
such as retail stores, supermarkets and
walkways.

Using video feeds, it allows estimations of


distances between people, and counts number
of instances when 6-ft distance (configurable
parameter) is not maintained between people.

Commercial in Confidence Detection of social distance index and counts of violations logged
Furniture and garbage detection
For the hospitality industry, we developed a
solution that would rate the cleanliness and
condition of a room provided a picture.

This included detecting the number of chairs in


a room (and whether they were occupied or
vacant) and the detection of garbage strewn
across surfaces (wrappers, torn labels etc.).

Technology stack used included Tensorflow


and YOLO.

Commercial in Confidence
AI Generated Sports Highlights
Producing sports events highlights is a
resource intensive, manual process today. Our
customer presented a novel challenge of auto
detection of game highlights using an ML
approach. We achieve this through

- Image and audio classification /


segmentation techniques e.g. analysis of
Video Analytics: Detection of Scoreboard from Video display
facial expressions, foul whistles, crowd
applause etc.
- The resulting system produces relevant
highlights measure up to viewer interest
and value added B2C engagement
methods through a customized app
Audio signal extraction Spectrograms on separated
Commercial in Confidence vocals/background
Live Game Score
Extraction
● Information Extraction e.g. timers,
scores, team names, event titles,
game halves, etc. from gameplay feed.

Commercial in Confidence
Auto Generation of
highlights from live
game
● Video shown on right is one
example of a highlight
generated
● Uses audio cheers /
commentator excitement / play
start and end
● Cheer meter on the left of the
video shows likelihood of the
moment being classified in the
highlight

Commercial in Confidence
Solar Systems feasibility using Satellite imagery

The System takes the address and energy


usage from user and uses this information to
determine the number of solar panels and their
placement on their particular house. We use
the address to get satellite image of their
house and use that satellite image to
determine:

1. Shape of their roof and other obstacles


on the roof to calculate the usable area
2. Direction of sun
3. Trees and other obstacles blocking the
sunlight from usable area of the roof.
Commercial in Confidence
Rank Sense
Detects different parts of a webpage through an image

Commercial in Confidence
Texture Synthesis
The solution achieves two main goals:
Inpainting and Texture Synthesis.

With inpainting, we fill unwanted parts of the (inputs)


image using the texture from the good parts.

With texture synthesis, we use small images


containing a certain texture and grow them
using generative models to create larger image
containing the same texture as in the small
image.

(output)
Commercial in Confidence
Ad Detection
A system that recognizes and separates ads, given an image containing many different
advertisements.

Commercial in Confidence
Surgical Instrument Detection in Operating Rooms

This system takes on a video as an input with surgeons using the


instruments and talking to their support staff. We have trained
two models:

● A model that listens to the audio and determines the name


of the instrument taken
● A model that looks at the images and determines the
instruments in view at the time

Since most of the instruments look very similar, it is very hard to


get good accuracy on image alone. So, we cross reference
detection from audio and video to give a final verdict.

Commercial in Confidence
Real-time Image
Detection of
Surgical
Instruments -
Demo

Commercial in Confidence
Real-time Image
Detection of
Surgical
Instruments - Demo

Commercial in Confidence
Real-time Image
Detection of
Surgical
Instruments -
Demo

Commercial in Confidence
Facial
Recognition for
attendance

Commercial in Confidence
Real-time depth
estimation

Commercial in Confidence
Computer Vision Skills & Strengths
METHODOLOGIES FRAMEWORKS & PLATFORMS
& ALGORITHMS LIBRARIES

Logistic Regression SkImage Dialogflow


XGBoost Tensorflow AzureML Studio
Gradient Boosting Keras Amazon Rekognition
RandomForest PyTorch Amazon SageMaker
GANs
RNNs
LSTMs
Residual/CNNs
Commercial in Confidence
We derive our strength from a diverse and
experienced team skilled across multiple Machine
Learning domains

How can we help you?


Engagement Process, Next steps

Review of Final project


Problem
initial results feasibility
Statement Investigation Production
outcome
Specification of Sample System
More data
Data Development
points as Validation &
Sample Data
inputs Acceptance

Estimated 2-4 weeks*

Commercial in Confidence

You might also like