AI, Computer Vision & Deep Learning Consulting

Big Vision LLC is a consulting firm with deep expertise in advanced Computer Vision and Machine Learning (CVML) research and development. We work on a wide variety of problems including image recognition, object detection and tracking, automatic document analysis, face detection and recognition, computational photography, augmented reality, 3D reconstruction, and medical image processing to name a few.

We provide turnkey solutions for clients who do not have the resources to build a complete CVML solution in-house. A typical CVML project often requires an extensive data collection effort, thoughtful dataset design for training and testing, research and development for implementing the state of the art solution and finally deploying a system that will easily integrate with the client’s existing systems. To deliver this end-to-end solution we employ a data collection team and a group of R&D experts and systems engineers who have experience in moving ideas from research to practice.

We have built scalable systems that are currently in production and are serving millions of requests a month in critical environments ( e.g. a large U.S. bank). Our solutions have a huge real-world impact in terms of speed and accuracy of algorithms, reduction in manual work, and identification and creation of intellectual property.

Computer Vision and Deep Learning Projects

In this section, we briefly describe a small collection of current and past projects. In some cases confidentiality agreement prohibits us from revealing details.

Image Recognition using Deep Learning

Our client, is the Netflix for legos. Users of their website receive a lego set for a fixed priced monthly subscription. After they have played with the lego set, they return it to receive a new set. The users are not charged for missing lego pieces but before sending out the same lego set to a different user, it needs to be ensured that the set is complete with all pieces. To accomplish this, our client had previously built an image recognition system to identify and count lego pieces with humans in the loop to ensure a very high degree of accuracy.

Our deep learning based image recognition system, that replaced their existing system, decreased human effort by over 70%.  We were able to further reduce their cost by taking over the remaining 30% of manual work.

Our training system can be used by people with no knowledge of computer vision. The new image recognition system continuously improves as humans in the loop re-train the system with hard examples on which the system fails. We integrated with their existing system via a robust web API.

Face Recognition

  1. Drones : We have built a deep learning based face recognition system that is deployed on a drone. We overcame several challenges to surpass our promised face recognition rate and speed. We ported our face recognition system onto an embedded device ( Jetson TX2 ) that was not yet fully supported by deep learning frameworks like Tensorflow. We also dealt with low resolution video data with severe compression artefacts, motion blur and lens distortion.
  2. Cell Phones: We worked with one of our clients for developing state of the art face recognition algorithms for mobile authentication. Our Deep Learning based approach significantly improved their recognition system that was based on traditional computer vision techniques. On extremely challenging examples under arbitrary lighting conditions, their recognition rate improved to over 95% with a false positive rate of 1 in 10,000. In addition, we also implemented spoof detection using Deep Learning which accurately detects spoof attacks based on a person’s picture.

Real Time Video Monitoring

We have built a deep learning based video analysis/monitoring/security system that dramatically decreased false alarm rate by using a combination of Artificial Intelligence and humans in the loop. The training system uses state of the art object detection and tracking algorithms and is trained using a large dataset collected by our data collection team for the client. The system is built using GPU instances on Amazon Web Services and is designed to handle thousands of camera feeds simultaneously.

ID Document Analysis

One of our clients is in the business of authenticating identity documents (driver’s license, passports, etc.) We have developed several imaging and image analysis algorithms that fall in the following broad categories —

  1. Geometric Correction: We have developed algorithms for automatically detecting and rectifying ID Type 1 (e.g. driver’s license, credit cards etc.), ID Type 2 (e.g. South African green book) and ID Type 3 (e.g. passports). Imaging these documents present significant challenges because of arbitrary camera orientation, glare, and bad lighting conditions.
  2. Fraud and Tamper Detection in digital IDs: We have developed Deep Learning based algorithms that automatically detect several different kinds of tampers in digital ID documents.
  3. Information Extraction using OCR: We are currently working on extracting text information from mobile phone photos of various id documents. The difficulty in this problem arises from images acquired by untrained users under uncontrolled lighting conditions that often include glare and blur. The images of a passport’s data page taken using a cell phone have significant curvature and this required us to solve challenging image alignment problems.
  4. Liveness Detection: Our video based liveness detection foils a large number of attack vectors with a very high degree of accuracy to ensure that the person being photographed for recognition is a live human.

Object detection in Aerial Images

We are currently working with a large sports analytics company on several projects that involve tracking the ball and players in video footage of a team sport. These problems are challenging because the setup is uncalibrated, and the imaging conditions are uncontrolled.

Video based Sports Analytics

We are currently working with a large sports analytics company on several projects that involve tracking the ball and players in video footage of a team sport. These problems are challenging because the setup is uncalibrated, and the imaging conditions are uncontrolled.
Photo-Enhancement.jpg

Photo Enhancement

One of our clients did photo shoots for real estate agencies. Earlier they were editing the photos manually to enhance the photos. It needed a big editing team, was subject to human errors and needed a long processing time for each property. We automated their pipeline to get the best results in the photos with minimal human intervention. We have delivered systems that enhance both the indoor and outdoor photos taken under varying lighting conditions and reduced the turnaround time drastically, while maintaining similar or even better quality images.

Shooter Identification

Working with a security company we developed a solution for identifying an active shooter in surveillance camera footage in a school environment. Compared to most object detection problems, this one requires an extremely low false positive rate and a very high detection rate for an extremely rare event (shooting).

Background subtraction

We implemented several custom algorithms that are being used in a medical device for accurately measuring the power of the corrective lens. Medical devices with a significant imaging component have very stringent algorithm requirements when it comes to accuracy and robustness. These algorithms involved object detection, real-time calibration and robust feature extraction.

Detecting parasites in Horse Feces

We implemented a Deep Learning based solution for locating and identifying different kinds of parasites in images of horse feces for MEP Equine Solutions.  The image on the right shows parasites Ascarides (A) and Strongyles (S) being detected while negatives marked as N.

Algorithms for a Medical Imaging

We implemented several custom algorithms that are being used in a medical device for accurately measuring the power of the corrective lens. Medical devices with a significant imaging component have very stringent algorithm requirements when it comes to accuracy and robustness. These algorithms involved object detection, real-time calibration and robust feature extraction.

Analysis of Microscopy Images

We have worked on the analysis of microscopy images to help focus a computer controlled microscope to the right location on a slide.

Vision Solutions for Retail Applications

We built a vision system that provides in-store analytics and retail shelf analysis for an IOT client.

Vision Solutions for Fashion

We have built three different applications in the fashion domain. The first project involved classification of products into more than 1000 different categories. The second project involved identifying fraudulent merchandize. The third is a visual search engine we built internally for a few fashion categories.

Visualization for Facial Cosmetic Surgery

We have built an application for simulating facial cosmetic surgery. The user will be able to visualize the effects of the surgery on their own photo. This required the precise localization of various facial landmarks on the face with very high accuracy

Facial Gestures

We are also working on a cyber security application based on facial gestures. This work is based on accurate tracking of facial landmark points.

3D Object Detection using Depth Images

We built a proof of concept system for a Fortune 500 company for detecting objects in rgbd ( depth ) images.