Improve the perceptual quality of an image through a deep learning model that does denoising, contrast correction, white balance correction, detail enhancement, etc.

Color orientation and exposure adaptive Deep learning based Low Light Enhancement algorithm for an output that looks natural

Exam Proctoring Application that tells whether the examinee is looking at the screen or not. The project was done as part of the capstone project of Plaksha Tech Leaders Fellowship, as Machine Learning Engineer Intern at Naukri.com (InfoEdge). 

Because of the lack of labeled supervised data to do the classification within 8 weeks, we used unconstrained face tracking, face recognition, head pose estimation, and gaze estimation models to achieve an accuracy of 92%. However, the screen resolution needs to be calibrated to correctly predict using the gaze angles.

Real time content aware 2x video super resolution that does motion artefact reduction and detail enhancement at 30fps @3.5TOPS

Pairing fonts is a classic issue in the world of design. Various fonts may be used to attract focus, to direct the eye, or even to shape the foundations of a brand identity. Instead of asking what is a good font pairing, which is a hard question to answer, let us focus on what are bad font pairings! Fonts that are very similar yet slightly differ lead to visual conflict, whereas totally different fonts are an obvious answer.

Instead of asking the designers what could be a similar fonts that goes well with another font, we can treat this as a vision task and visualize the cluster of all fonts.
  


This project extracts features of fonts from font images and uses a modified distance/similarity function, that can effectively determine similar, dissimilar, and pleasingly contrast fonts.

Understanding where people look and click on visual designs can provide clues as to how the designs are received and where the content is most appropriate or relevant. This is crucial in determining/quantifying an observer's attention to evaluate visual designs.

The project uses neural networks trained on user clicks and importance annotations on hundreds of designs. paper

Human gait, or a distinct walking pattern, is a biometric pattern which allows identification of an individual when other biometric features like face or iris is not visible. With the existing state-of-the-art models, we can’t handle inter-distinctive inter-class differences of human gait in real-time. In this project, we proposed a novel approach called GaitNet which directly learns a mapping from the pose extracted images to compact gait embeddings in the semantic space where distances directly correspond to a measure of gait similarity. On top of it, we have developed a Siamese based neural network gait recognition pipeline to extract robust and discriminative gait features for human identification. Our experiments on the gait recognition framework demonstrate the efficacy of the approach. 

The project was implemented by a team of five Plaksha Tech Leaders Fellows, as a deliverable for Data X-1 Lab. 

Guides:
1. Alexander Fred-Ojala, SCET, UC Berkely
2. LinkProf. Ravi Kothari

This project uses Pytorch to perform Object detection on Kitti Dataset. Faster-RCNN is used with Resnet-50 backbone having feature pyramid network.

Performance : [ mAP (0.50) = 0.944 ]

This project uses Keras based Yolo-v3, trained on Pascal VOC-2012 dataset.

Key contribution : Improved performance and accuracy optimization using Online Hard Example Mining (OHEM). Reference

Performance : [mAP = 0.75]

Built the backend of a Face and Expression Recognition App as part of a team of three for the final project of advanced AI & ML certification at IIIT-H. The app used a Pytorch based Facial Expression recognition system using the CK+ dataset on Mobilenet v2.

Six classes {Anger, Disgust, Fear, Happiness, Sad, Surprise, Neutral}

Performance : [Accuracy = 94.64]

Award :  🥇 Secured First rank in the capstone challenge and was awarded the certificate with Outstanding distinction.

This project experiments with the choice of Style Image for effective Neural Style Transfer. The results show that the Content and Style Images need to be semantically similar for the pre-trained weights to effectively generate a style transferred image.

This project was done as a research project at Cellstrat AI Labs. A Cellstrat webinar explaining this.

This project was part of the second hackathon for the advanced certification in AI & ML at IIIT-H. The purpose was to make a speech based menu for a restaurant. The application accepted voice commands from 0 through 9 for items and amount. An LSTM based network was used to classify input audio as a numerical value. It mainly utilized the MFCC vector as an input to a multi-layered LSTM network.

PreviousNext

The idea we were working on was Grammarly for visual design - Design Validation using AI.

Although there were numerous learnings, my biggest takeaway would be that 'A problem or a potential idea, no matter how good it seems, stays only a hypothesis unless your MVP is verified by your potential customers. '

Detailed Blogpost on the event : First Cellstrat AI Hackathon (Computer Vision)

PreviousNext

Cellstrat AI Labs weekend meetups happen every saturday in Bangalore, where researchers and AI practitioners meet to discuss latest research papers and showcase applications.

Some blog posts about a handful of meetups:
1. 2nd March 2019
2. 9th March 2019
3. 23rd March 2019

Webinar Link : Effect of choice of Content-Style Image Pair on Neural Style Transfer

A Case Study on Healthy Employee Initiatives in corporates like Nagarro.

Problem Statement : "While Nagarro aspires to become ‘Most Healthy Workplace’ plus wants to create Social Impact, its employees generally lack the motivation to maintain a healthy lifestyle."

User interviews were conducted to connect the dots between what we knew and what wedidn’t know.

Guide : Prof. Jitesh Panchal, Purdue University

Computer vision technology is becoming very popular and its application ranges from surveillance, retail, the agriculture and food industry, social credit system, etc. The technological advancements such as integrated hardware components, high computation power, deep learning and advent of data have made it possible to use the technology in real-time. With all these applications, there are conflicts of values between all the stakeholders, and it raises ethical questions.

This case study  examines the ethical aspects of video surveillance and it leading to a social credit score, from the perspective of various stakeholders.

Guide : Prof. Andrew O. Brightman, Purdue University