Working on my thesis around developing and optimizing the traffic analytics system on edge for highways.

Teaching assistant for the course EECS2101 - Fundamentals of Data Structures

Teaching Assistant for the course EECS1015: Introduction to Computer Programming

Teaching assistant of EECS1720 Building Interactive Systems

Teaching Assistant for the course EECS 1015 - Introduction to Computer Programming.

A platform to share the best upcoming research ideas in the field of AI.With an advent of the vast research in recent years, young researchers, developers, and MNCs are targeting the AI as an element for their startups, projects and introducing novelties. Due to the drastic drift of the research, more and more new and improved research papers are published everyday which hinders the pace of the R/D work and to get the best out of it. Specific sets of problems requires fine- grained functions for completion that are very hard to find directly from the research papers due to their redundant and recursive nature. We developed a platform, so that any researcher can share their views on the upcoming work/novelty in AI by following specific set of guidelines to complete the seamless transition from research to deployment. For contributing to our blog, reach out to us at team@visionwizard.in

Research and Development for optimizing Convolutional Neural Networks1. Site Specific Training Tool: Designed and developed a proprietary light classifier architecture from scratch by incorporating the present and efficient in-house model backbones of the object detectors which helped us to increase the mAP by 8% on the given site. Also integrated the mod- ified open-source Background Subtraction algorithm and introduced update-detect framework in this tool which assisted to get the seamless results on the low-FPS video streams of the surveillance cameras.2. Low Compute - High FPS BGS Algorithm: Developed an end to end framework of the low com- pute background subtraction algorithm for videos viz. Grid Temporal Median running at 70 FPS on Intel low-powered devices with the pruned classifiers for best performance. This pipeline also helped in detecting and correctly classifying objects when Deep Learning based Object Detectors failed steadily on difficult scenes such as Fish eye Cameras, noisy, and Non-IR fields of vision3. Redefining One-Shot Object Detectors for Two-Class Problem: Developed an hierarchical clus- tering approach for the anchors of One-Shot Object Detectors namely Anchor Search Algorithm for assigning and learning dynamic optimal anchors for two different classes individually . 3% mAP gains were observed after final bench-marking the improvements.4. BGS Pipeline: Leveraged the traditional Computer Vision based background subtraction algo- rithms with a very lightweight object classifier with the new Update Background Policy for blob detection on low-powered devices.

Made an end-to-end framework for optimization of the One-Stage Object Detectors for em- bedded devices. The pipeline included the below given sub-stages to accomplish the same.1. Network Pruning by conniving the algorithm for handling the residual chains present in the back- bone architectures for halting recursive chain removals. This standalone stage provided with 2-3X speedup and 3-4X memory improvements on embedded devices.2. Quantization of object detectors by leveraging dynamic ReLU activation function for INT-8 calibration and quantizing the weights of the model with the help of different compression algorithms (standalone implementation). Dynamic ReLU viz. QReLU helped us to get only 0.5% drop after quantization and 1.2% increase in the overall mAP during training than the baseline.