Danish Ansari

Stats Perform

AI Scientist

2021-current

Working on developing AI models for live game analysis. Working with Players are challenging as all players wear similar jerseys and thus extracting individual players information becomes difficult.

Synapsica Technologies

AI Scientist

2019-2021

Worked on MRI-scans to detect multi-label, multi-class disease on intervertebral spinal discs, extracting underlying features through CNN. Got familiarity with medical data format and working principles. My role involved doing RnD for product, developing neural architecture, training model, optimizing model, deployment and mentoring team.

Graymatics

AI Engineer

2018-2019

As an Engineer at Graymatics, I enjoyed working and exploring the way CNN’s can be tweaked to achieve desired goals. Learned Python frameworks like Tensorflow and Keras. During my work I developed an end to end ANPR system with a reasonable accuracy for CCTV cameras.

Player Tracking - Soccer

This is one of the most challenging tasks in computer vision I came across. In most games players wear very similar jerseys. And in games like soccer or basketball, it is really difficult to track a player of interest. I looked into many solutions that exist and understood it’s ideas, most of which are for crowd human tracking or vehicle tracking. And I am working on deriving appearance based embedding learning to differentiate between players.

Player Reid (clustering) - Soccer

Player reid is yet another challenging task, where we want to cluster players from different teams and responsibility into separate categories. I explored different approaches and came to a conclusion that similarity matching can be a good strategy as players from one team appear similar. I used CNN and Triplet loss to make the model learn to detect players from the same teams and then used a clustering algorithm to group them.

Spindle X

Working with Xray images comes with an inbuilt challenge of data variation due to no standard rule to take pictures. With huge augmentation we wanted to detect each corner point in the spinal intervertebral discs. Because these fine points can be used to further measurements and make the diagnosis faster and better.

Spindle MRI

Worked on detecting disease at disk level in an MRI Image. So it was a multi-label multi-class problem. Used CNN to get underlying features to create a one vs all classifier. Detection of such disease needs very fine grained features, as it can be a difficult task even for human expertise.

ANPR (automaic number plate recognition)

Automatic Number Plate Recognition is one interesting challenge which can have many uses and at a large scale. Detection of the correct licence plate can be used for surveillance as well monitoring any potential traffic violations. Combining CNN and RNN to form RCNN was the key ingredient. Each row of the plate was passed a sequence and it was able to predict what the potential characters could be even if the characters are not fully visible.

Helmet Violation Detection

Convolutional Detection methods like YOLO, SSD come with computation cost. So I Converted this problem into classification. I used license plates position to estimate potential head location and did a three classification as if both rider and pillion is wearing a helmet or not or if the image is junk since we don’t want to make false predictions for people who are actually following traffic rules. Having this in mind I was able to get about 98% recall.

Vehicle Detection & Classification

Due to high computation cost, using convolutional detectors was not an option, but with the information of the plate we could make estimate about the vehicle presence and whereabout, did a classification and optimized it to run on intel machine at real time.