Useful Links: Github / LinkedIn / Google Scholar
I'm currently a PhD Student in the Elmore Family School of Electrical and Computer Engineering, Purdue University. I am also a Graduate Research Assistant affiliated to the Video Analytics for Daily Living (VADL) lab. Previously, I was a research assistant at the Indian Institute of Science
PhD in Electrical and Computer Engineering, (Communications, Networking, Signal & Image Processing)
- Performed a systematic analysis of face image quality evaluators such as SER-FIQ and SDD-FIQA.
- Leveraged specific quality estimation and datasets for a deeper understanding of learning-based quality estimators.
- Employed prevalent distortions in face images to develop superior and more robust face quality estimators.
- Introduced a new, operational evaluation protocol that minimizes the computational complexity of assessing face image quality estimators.
- Devised a task-oriented CU Frame Partitioning procedure for video encoders like HM and VVC.
- Employed lightweight, edge-based neural networks that predict frame partitioning depending on the task to systematically aid the encoder i.e., for region-specific video encoding.
- Achieved bit-rate conservation during transmission and diminished encoding time while ensuring the performance of learning-based analytics remains.
- Unifying the representation of audio and video with a singular Neural Field Representation.
- Worked on methods of Model Pruning and Quantization instead of focusing on modality-specific compression.
- Investigated the impact of compression on computer vision tasks such as pedestrian detection and face recognition.
- Assessed compression's influence on task performance in a variety of conditions, including differing light, resolution, camera models (e.g., fisheye), camera streams (such as RGB vs IR), facial skin tones, object dimensions, etc.
- Appraised the impact of various encoders and their configurations on task performance.
- Pre-processing real-world data, eliminating bias in terms of evaluation scenarios to create interpretable results
- Creating consistent annotations for fair performance evaluation
- Exploring the performance co-dependence that exists between face detection and recognition tasks
- Using multi-view, multi-modal data to help with end-to-end detection and recognition in different illumination conditions and recording environments
- Encoding neural network features using existing vidoe codecs to see if it is a better alternative to encoding images/frames
- Working to understand if neural networks can be split effectively such that intermediate features can be encoded and transmitted
- Finding network splits such that the best balance between saving bit-rate and maintaining task accuracy can be obtained
- Developing Feature-To-Image mapping \& Auto-Encoder architectures for dimensionality reduction of features before encoding
- Developed a Robust PCA based saliency predictor that helps with the background-foreground segmentation,identificationand population estimation of fauna in camera-trap images from Senegal
- The RPCA method does not require training and it’s performance is comparable to learning models like R3-Net
- The saliency predictor is used to detect and track animals, estimate population density and regular animal activity patterns
- Worked with Dr. Reshmi Mitra from Southeast Missouri State University in developing a novel framework that helped in detecting DDoS attacks on Edge Devices using Recurrent Neural Networks
- Achieved SOTA performance on the UNSW 2015 dataset while ensuring minimal model architecture i.e can be runon edge devices
- Worked with Dr. Abhay Sharma and Dr. Raghu Krishnapuram in developing solutions to help in Traffic Analyticsthat involved tasks like vehicle counting, license-plate detection speed estimation and queue-length estimation
- Built a real-time front-end WebServer System that serves live video streams in RTMP and HLS formats. Implementedthe server with methods for Discovery,Sharing Content,Routing,Congestion and Load Balancing
- The entire framework has been deployed in Electronic City, Bangalore, India
- To allow for simpler marker insertion into H.264 video streams that can not only hold additional metadata such as sensor values corresponding to frames, bounding box values etc but can also help in live stream synchronisation.
- An understanding of whether second-order Quadi-Newton methods can be used for practical Deep Learning applications. If so, can they compete with popular first-order methods like SGD.