Research Experience

• ADVISOR           : Prof. Vijay Ganesh

• PROJECT TITLE : Logic-based Automated Program Translation

Developing a NLP- and logic-based automatic program translation (e.g. Java to Python) system to generate compilable target source code and ensure complete equivalence in runtime behavior

• ADVISORS         : Prof. Niloy Ganguly, IIT Kharagpur;

                              Prof. Krishna P. Gummadi, Max Planck Institute SWS, Germany

• COLLABORATORS : Mr. Gourab K. Patro, IIT Kharagpur;
      Dr. Abhijnan Chakraborty, IIT Delhi

• PROJECT TITLE : Fairness in Virtual Conference Scheduling 

For virtual conference scheduling, it is a challenging feat in accommodating participants from different time-zones as well as accounting for their interests in attending various talks. To address this, we have developed a joint optimization framework whereby, conference organizers can design talk-schedules that balances (i) efficiency objective of maximizing participation and (ii) fairness objectives of stakeholders (participants and speakers) from various time-zones. 

• ADVISOR            : Prof. Niloy Ganguly (Visiting Professor at Leibniz Universität Hannover, Germany; Professor at IIT Kharagpur)

• COLLABORATOR : Prof. Romit Roy Choudhury, University of Illinois, Urbana-Champaign

• PROJECT AREA   : Ethical Machine Learning, Opinion Dynamics & NLP

Ethical machine learning aims to develop machine learning methods whose outcomes do not have a disproportionally large adverse impact on particular groups of people sharing certain sensitive traits such as race or gender. Through this project, we would design human-centric ML models through counterfactual fairness, to tackle such bias of algorithmic decision-making. We are experimenting with opinion dynamics, where the personal beliefs of individuals get influenced through interactions and dominations. Using game theory, reinforcement learning, and opinion mining, we estimate the likelihood of polarization in the population.

• ADVISORS        : Dr. Abir Das, Assistant Professor, IIT Kharagpur;

                             Dr. Rameswar Panda, Research Staff Member, MIT-IBM Watson AI Lab

• PROJECT AREA : Video Action Recognition & Zero-Shot Temporal Detection 

Worked on simultaneous recognition and localization of activities in untrimmed videos, on the time-axis. The project would address the scarcity of labeled data to detect activities that have never been seen during training phase. Through this project, we would explore the Region Convolutional 3D Network (R-C3D) architecture and aim to design a zero-shot end-to-end variant for the same. 

• ADVISOR          : Prof. Nilanjan Ray

• PROJECT AREA : 3D Object Detection & Pose Estimation in Indoor Scenes 

3D object detection in indoor scenes from RGB-D information and point clouds. In this project, we aim to formulate an end-to-end trainable neural network that detects amodal 3D object bounding boxes from a 3D volumetric scene.

• ADVISOR           : Prof. Sanjoy Kumar Saha

• PROJECT TITLE : A Real-Time Semantic Image Segmentation Approach Through Region-Partitioning

Developed a fast semantic image segmentation system, that is suitable to be applied in real-time robot vision and semantic SLAM. Firstly, we have compared and contrasted two statistical region-partitioning techniques that partition a natural image into numerous superpixels. For classification, two multi-stream ConvNet architectures were designed, that possess a high receptive field and are capable of maneuvering different neighborhood contextual information (Field-of-Views). By this method, we could bypass the heavy computation involved in a dense pixel-wise labeling of Fully Convolutional Network (FCN).

• ADVISOR            : Dr. Partha Pratim Mohanta, Associate Scientist

• PROJECT TITLE  : Spatio-Temporal CNN-RNN Decision Fusion Strategy for Event Classification in Unconstrained Videos

Developed a multi-tier late decision fusion technique that utilizes “biased conflation” of probability distributions, to consolidate frame- and video-wise event predictions from CNN and RNN respectively. For this, two types of probability distribution aggregation strategies were conceptualized: Cross-Fusion and Self-Fusion. Ultimately, decisions were aggregated in a hierarchical and multi-tier scheme. Further extended this work, by spatio-temporal action localization to recognize high-motion patches in key-frames that assist the classification task.

• ADVISOR           : Dr. Partha Pratim Mohanta, Associate Scientist & Summer School Coordinator

• PROJECT TITLE : Key-Frame based Event Recognition using Temporal Features & Deep Residual Networks

Aiming temporal localization of actions, a video was symbolized by a number of representative (key) frames. First, temporally redundant information was identified and discarded after quantifying 'motion' encompassing a frame through motion histograms. Next, a graph-based key-frame selection technique is used that involves choosing a set of most temporally distinct frames, alongside maintaining a lower limit of difference in timestamps. Finally, a hybrid model consisting of deep residual network and LSTM captures both the short- and long-term temporal motion patterns, and ultimately predict event for a video. 

• ADVISOR           : Dr. Debrup Chakraborty, Associate Professor and Head (CSRU Dept.)

• PROJECT TITLE : Carry-less Karatsuba over Galois Field in Block Ciphers

Using a processor instruction for 64-bit carry-less multiplication, efficient “carryless Karatsuba multiplication” was implemented to multiply degree-127 binary polynomials. Subsequently, efficient reduction of the degree-254 binary polynomial, modulo a specific pentanomial (defining finite field of GCM) was executed. The work served as a base for using AES-GCM in high-performance symmetric key block ciphers. During this project, substantial experience was gained in coding with Intel Intrinsics API.

• ADVISOR            : Dr. Ram Sarkar, Associate Professor

• PROJECT TITLE  : Document Analysis & Medical Image Segmentation

• SUB-PROJECTS   :

  • co-supervised by Prof. Mita Nasipuri --> Handwritten Document Image Binarization for degraded historic documents & Handwritten and Printed Character Recognition for English, Bengali and Telegu

  • co-supervised by Prof. Subhadip Basu  --> Binarization of 3D MIP confocal microscopic images of Hippocampus area of mouse’s brain

Estimated background of degraded historical documents by a novel ProximBackFlood technique that carries out inpainting of nearest background pixels in foreground areas; optimal search by generalization of Fermat's Theorem. Segmentation was experimented with K-Means, Fuzzy C-Means Clustering and histogram thresholding on adaptive image partitions. This was subsequently utilized in OCR, where a tailor-made A* algorithm was developed to construct character skeletons. Developed an improvised version to binarize confocal microscopic brain images.