Sampanna Kahu
About me
An aspiring researcher with interests in Computer Vision and Natural Language Processing.
Presently working at Google (California).
Thesis title: Extracting figures and tables from scanned Electronics Thesis and Dissertations (ETD).
Using Deep Learning (Resnet-101) to segment figures and tables from the PDF files of scanned ETDs to aid in downstream information
retrieval tasks.
Thesis document
Project images
My projects
| Sr. No. | Name | Description | Worked with |
|---|---|---|---|
| 0 | Deep task scheduler for Linux Kernel | (Linux kernel, Deep Learning, RNN) Using an artificial neural network to schedule tasks in Linux kernel. This project was undertaken as part of the course ECE-5984 Linux Kernel Programming at Virginia Tech in Fall 2019 instructed by Prof Changwoo Min. Full report (PDF) | Prof Changwoo Min |
| 1 | Few shot distillation learning using Monte Carlo Dropout | (Bayesian Deep Learning) Used Monte Carlo Dropout for enabling few-shot knowledge distillation from a teacher model into a student model. Image 1 Image 2 | Myself |
| 2 | Trajectory prediction of pedestrians using Monte Carlo Dropout for autonomous cars | (Bayesian Deep Learning) Trained an RNN (LSTM) to not only predict the trajectory of a pedestrian, but also to predict the probability distribution of possible trajectories using Monte Carlo dropout. Image 1 Image 2 | Dr. Prahlad G. Menon |
| 3 | Fluoroscopy based sparse reconstruction of coronary venous anatomy with application to cardiac resynchronisation therapy | We worked on the fluoroscopy-based sparse 3D reconstruction of coronary venous anatomy with application to cardiac resynchronisation therapy (CRT). Using this technique, a better informed placement of the CRT pacing leads can be made during a cardiac surgery. The image processing pipeline in this project uses ridge detection to identify a wireframe structure of the coronary venous anatomy from projection images followed by tomographic reconstruction to obtain a 3D model of the coronary arteries. Image 1 Image 2 Image 3 | Dr. Pratap Tokekar |
| 4 | Detection of Diabetic Retinopathy from retinal images | Developed various image processing techniques to develop an algorithm for automatically detecting and measuring the severity of Diabetic Retinopathy from high resolution retinal (fundus) images. | Dr. K. M. M. Rao |
| 5 | Image segmentation in pure Java (without libraries) | Implemented an object oriented program in Java to segment a black and white image to extract out contiguous clusters of dark pixels. Source Code | Myself |
| 6 | ML Challenge | Participated in a competitive coding Machine Learning challenge organised at Flipkart.com. The objective was to detect the return-to-origin fraud happening at flipkart. Source code on Source Code | Myself |
| 7 | QNA Classifier | Made a Deep Learning sentence classification ML model using TensorFlow. This model classifies any input sentence into a question or an answer. Source Code See Demo | Myself |
| 8 | Parallax Error Lens | Parallax Error Lens can convert any ordinary surface into a touch-surface using stereoscopic vision principles. It is a specially designed lens which can be attached to a camera to give it depth sensing capabilities. Note: (This project was selected for showcase in BITSAA Global Meet 2014 and has won third prize in Idea Contest organised by Centre for Innovation, Incubation and Entrepreneurship (CIIE) at BITS Hyderabad). | Myself |
| 9 | Robotic arm | A 2-DOF planar manipulator arm that uses two servo motors to mirror cursor movements. Source Code | Myself |
| 10 | Android app to download a random wallpaper from the internet | Made an Android app, which when triggered, downloads a random wallpaper from the Internet using Bing's API and sets it as your Android smartphone's wallpaper. Source Code | Myself |
My awards and achievements
| Sr. No. | Name | Description |
|---|---|---|
| 1 | Automation and Robotics Club | Secretary and founding member of Automation and Robotics Club (ARC) at BITS Pilani |
| 2 | 1st prize in Computer Vision and Robotics competition | Position 1st in Machinist (Real-time video processing and robotics competition) BITS Pilani National Technical Festival (ATMOS '13). |
| 3 | 1st prize in Robotics competition | Position 1st in Maze Solver Competition at BITS Pilani National Technical festival (ATMOS '12). |
| 4 | 3rd prize in Robotics competition | Position 3rd in the zonal of International Robotics Challenge Gridmasters Competition at IIT BOMBAY Techfest |
| 5 | iPodia Global Innovation Conference at Beijing, China (May 2015) ( www.ipodia.usc.edu ) | Participated in the iPodia innovation meet-up organised by University of Southern California across 5 nations to open up new innovation opportunities by leveraging cultural diversities. |
| 6 | Robotics Mentor (2013 - 2014): (Student Mentorship Program at BPHC) | Mentored students junior to me in robotics to help them get familiar with related concepts and helped them participate in competitions |
| 7 | Phoenix Office bearer (2013 - 2014) | Was a core member in the student committee of the EEE and ECE technical association at BITS Hyderabad. Helped organise various technical activities and competitions |
My open-source contributions
| Sr. No. | Name | Description |
|---|---|---|
| 1 | Escape HTML only once
(contributed to Apache Commons Text). Source Code |
If already escaped HTML characters are in the input text, StringEscapeUtils.escapeHtml4() escapes them again. For example if the input is: 100kg & lt; 1000kg Then the output of escapeHtml4() becomes: 100kg && amp;lt; 1000kg At my workplace, we felt the need for a method in StringEscapeUtils which does not escape already escaped characters. |
My Hobbies
In my free time, I love to play the Tabla.
Tabla is an Indian percussion musical instrument originating in the 18th century.
