Human Pose Estimation with Domain adaptation

Company: Samsung Research America

Lab: Think Tank Team

Location: Mountain View, CA

role: Intern

My project was focused on developing dataset and network for estimating human pose from an untraditional sensor. From a high-level perspective, I developed domain adaptation algorithms to leverage existing data. I also developed pose estimation network and trained the same.

Ground plane detection using homography

Company: Texas Instruments

Team: ADAS SW

Location: Bangalore, India

role: Intern

My task was to detect ground plane from monocular images. I used homography as a tool for plane detection. I improved the technique by making the ground plane the dominant plane by developing an algorithm that chooses an ROI where the dominant plane is the ground plane. Since the motion is not exactly a rotation, the algorithm takes care of patches on the road such as the lane markings.

Weakly Supervised Object Detection

role: author

Replicated the papers Is object localization for free? and WSDDN for understanding weak supervision based object detection

Intersection reconstruction from single stationary camera

role: author

Worked on reconstructing intersection from a single stationary camera using a combination of Deep Learning techniques, non-linear optimization and tracking

Object tracking

role: author

Implemented Inverse Compositional Lucas Kanade algorithm with basis to perform object tracking.

Digital art using SfM

role: author

Created a tool to perform background blur, background replacement using stereo images or consecutive images in video.

Project website

Intelligent Inpainting

role: author

Intelligent Inpainting is a Computer Vision application developed in C++. It takes a single image and a user click. The click is intented to select a person on the input image. Then the application intelligently crops that person out of the image by filling with the background.

Project website Github Repository

Structure from Motion

role: author

Developed a SfM application in C++ from scratch. The project provided good opportunity to understand multi-view geometry.

Realtime sudoku solver

role: author

An android application that uses camera to search for sudoku in its view and solves it upon detecting one.

Playstore link

Astro Adventure

role: author

Trion the shape shifting alien has fallen from his spaceship. Help him get back to the ship by swiping your way around the galaxy. Try and keep him within the screen. If he falls out of the screen, he's dead. So swipe wisely!!! And have fun.!

Playstore link

Learning hierarchical policies in dynamic environments

role: author

This project is focused on developing an algorithm that quickly adapts to a dynamic environment with sparse reward structure. While a sparse reward environment is a quite challenging setting, a combination with a dynamic environment makes it even more challenging and such a setting is realistic too.

Shyam Jagannathan, Kumar Desappan, Pramod Swami, Manu Mathew, Soyeb Nagori, Kedar Chitnis, Yogesh Marathe, Suriya Narayanan and Anshu Jain:
Efficient object detection and classification on low power embedded systems
2017 IEEE International Conference on Consumer Electronics (ICCE)

Suriya Narayanan and Nagendra Gulur:
Understanding the performance benefit of asynchronous data transfers in OpenCL programs executing on media processors
IEEE 22nd International Conference on High Performance Computing (HiPC), 2015.

Suriya Narayanan and Prashanth Viswanath:
Improved Ground Plane detection in real time systems using homography
ICCE 2014

Sandeep Nital David, Sachin P Santosh and Suriya Narayanan:
Control Strategy for power flow management in a PV system supplying electrical loads in an air-conditioned bus
IEEE International Conference on Power Signals Control and Computations (EPSCICON), 2014.

Suriya Narayanan and Prashanth Viswanath:
Ground plane detection US 20150178573 A1

Action recognition

The idea of this pet project is to experiment with the effectiveness of LSTM and GRU for action recognition over averaged classification score (averaged over a certain set of frames).

Deep Q Networks

Coded various Q networks such as linear Q, non-linear Q, deep Q networks to train an agent for playing atari games

Advantage Actor Critic based RL

Coded A2C algorithm to train an agent for playing atari games

Augmented Reality

Coded an Augmented Reality application from scratch in Matlab.

Panorama

Create an application for two image image stitching in Matlab from scratch.

Scene Classification

Code Spatial Pyramid Matching from scratch in Matlab for scene classificaiton.

Decision Tree ID3

Developed the ID3 training algorithm in python from scratch.

Fully Connected Network

Developed a fully connected network in python from scratch. The code contained portions for feed-forward propogation and backward propogation.

Eye Mouse

Developed a computer vision application that tracks the eye ball movement and controls the mouse pointer accordingly. OpenCV 2.0 was used for Image Processing and Qt was used for user interface development.