Abhinav Narayan Harish
I am a CS PhD student at University of Wisconsin Madison working with Professor Josiah Hanna. Prior to this, I was a Master's student with Professor Zsolt Kira working on using Reinforcement Learning (RL) for long-horizon home-rearrangement in Habitat2.0. I received my Bachelor's degree in Electrical Engineering with a minor in Computer Science from IITGN .
CV / 
Email  / 
Google Scholar  / 
LinkedIn  / 
GitHub  / 
Twitter
|
|
Research
Broadly, I'm interested in Reinforcement Learning (RL) and it's application in the field of Robotics. My goal is to explore the limits of generalization in long-horizon RL across tasks and applied to distinct environments. To this end, I have explored hierarchical RL using Skill-Chaining for my Master's thesis. With a background in 3D Computer Vision, I'm interested in leveraging this to bridging the Sim2Real gap, enabling seamless interaction in the real world.
|
|
Skill-Chaining for Long-Horizon Rearrangement
Abhinav Narayan Harish,
Master's thesis, Georgia Institute of Technology
pdf /
Given a set of pre-trained skills performing specific aspects of rearrangement , i.e Navigating, Picking up or Placing objects, we developed a hierarchical fine-tuning scheme that can fine-tune these policies simultaneously. Our method addresses the hand-off challenge in rearrangement where subsequent skills are not aligned (i.e the navigation skill-terminates too far from the couch). Our method demonstrates superior performance (about 13% in rearrangement success) over a static policy.
|
|
RGL-NET: A Recurrent Graph Learning Framework for Progressive Part Assembly
Abhinav Narayan Harish, Rajendra Nagar,
Shanmuganathan Raman
Winter Conference on Applications of Computer Vision
(WACV 2022)
pdf /
bibtex
We propose an assembly framework that can assemble a shape in a canonical order by progressively gathering information. Compared to prior frameworks, our method achieves upto 10% improvement in part accuracy and 15% improvement in connectivity accuracy.
|
|
Double Compression Detection of Distinguishable blocks in JPEG images compressed with the same quantization matrix.
Abhinav Narayan Harish*, Vinay Verma*,
Nitin Khanna
International Workshop on Machine Learning for Signal Processing
(MLSP)
,
2020
pdf /
bibtex
In this paper, we propose a deep learning based approach to localize forgery in a JPEG image. We develop a multi-coloumn CNN architecture that utilizes spatial and frequency domain information and classify each 8x8 JPEG block as single or double compressed.
|
|
Neural Network based Block-Level Detection of Same Quality Factor Double JPEG Compression
A.Deshpande*, Abhinav Narayan Harish*, S.Singh*, V. Verma,
Nitin Khanna
International Conference on Signal Processing and Integrated Networks
(SPIN)
,
2020
pdf /
bibtex
We detect double compression at the patch level (64 x 64) or (128 x 128) by introducing an additional feature based on difference of DCT coefficients. Our classification network achieves upto 1.52% improvement in detection accuracy.
|
|
Language as a Means of Shape Differentiations
Statistical Machine Learning, (ECE-6254) Georgia Tech
pdf
We utilize language tokens to understand distinctions between shapes. Given a text-label we find the closest match to the target shape.
|
|
Inverse Reinforcement Learning on Multi-Armed Bandits
Online Decision Making in Machine Learning (ECE-8803), Georgia Tech
pdf
We study the problem of Inverse Reinforcement Learning on the LinUCB algorithm and obtaining an optimal regret gurantee.
|
|
Eulerian Motion Magnification
3D Computer Vision, IIT Gandhinagar
code
We develop a python implementation of the Eulerian Motion Magnification algorithm for revealing subtle changes in the world.
|
|
OpenGL Buzzy's Bowl
Advanced Programming Techniques (ECE-6122), Georgia Tech
We simulate the motion of UAV's as they are launched from a football field and move randomly on a sphere colliding with each other.
|
|
Interactive Interpolation between Splines
Computer Animation (CS-7496) Georgia Tech
code
A tool for interactively interpolating between points using 4 different spline types.
|
|