Susmit Agrawal

Advisor: Prof. Vineeth N. Balasubramanian

Relevant Courses: Machine Learning, Data Warehousing and Data Mining, Data Structures and Algorithms, Digital Image Processing, Natural Language Processing

Performed research on tuning pre-trained unconditional GAN models for novel view synthesis, with the help of synthetic data (in collaboration with Dr. Varun Jampani).
Investigated self-supervised methods for digital image forensics (in collaboration with Dr. Maneesh Singh and Dr. Toufiq Parag).
Developed a self-supervised method for keypoint detection (in collaboration with Dr. Varun Jampani) with knowledge distillation.
Worked on an RNN based approach for HDR deghosting with variable lengths of input image sequences.
Developed a few-shot HDR deghosting method with pseudo-labelling through weak self-supervision.

Developed a method for HDR deghosting at high resolution on limited hardware.
Developed a motion-segmentation guided fusion network for HDR deghosting without requiring optical flow correction.
Developed an augmentation based approach for HDR deghosting with static sequences.

Worked on a method to optimize the amount of water delivered to a crop through a drip irrigation system.
Developed a medical question-answering system using BERT and GPT-2, using anchor representations to generate repeatable answers for different versions of a given question.

Working on building Computer Vision models for Image Super Resolution.
Working on model architecture optimization for deployment in restricted environments.
Working on finding and fixing failure cases of trained models based on performance on real-world validation sets.
Working with traditional image processing algorithms for picture quality enhancement.

Worked on developing differentiable mathematical models to model propeller designs.
Developed testing rig software to log propeller performance along a range of metrics.

Worked on the implementation of CNNs and GAN architectures with resource constraints.
Finished Top 5 (among 1072 participants) in the in-house Stanford Tiny ImageNet Challenge.
Built high-accuracy classifiers for multiple datasets with constrained hardware and limited number of parameters.
Implemented Face Ageing using Conditional CycleGANs on the UTK face dataset on the cloud.

Worked on YOLOv3 for license plate detection in a video feed.
Deployed models of AWS EC2 instances, edge devices, and Android using Tensorflow Lite.
Developed architectures combining CNN and RNN model for optical character recognition.

Worked on the "Giggle" app for the Indian market.
Used Google Cloud's Computer Vision APIs to implement certain features.
Worked on image processing on raw bitmaps.