Projects

Trajectory prediction of autonomous agents
Honda Research Institute USA
  • Future trajectory prediction for all agents in a scene in a single shot in constant time. Single shot nature of this work makes it faster than other works in crowded situations (published at IROS 2020). Single shot nature achieved by predicting composite fields from neural network. Non-Local interaction block used to capture interactions between agents. Semantic segmentation features used to make semantically aware predictions.
See video of IROS 2020 presentation below!

SSP: Single Shot Future Trajectory Prediction



Scene Classification of road scenes 
Honda Research Institute USA
  • Worked on road place and condition classification on videos obtained from a front facing camera mounted on car. Place classification includes classification of road scenes, road surface conditions, road weather and road type (paper published in ICRA 2019). 2 stage neural network used – first stage selects candidate video snippet from long video, second stage classifies the snippet.
  • Improved Road place classification performance by redesigning neural network architecture and adding semantic segmentation as auxiliary task. F-score improved from 28% to 40%. 
  • NEW : Dataset webpage

Active Interpretation of Disparate Alternatives (AIDA) Project
Prof. Shih-Fu Chang, Columbia University
  • Worked on the AIDA project with Prof. Shih-Fu Chang in the DVMM lab.
  • My work involved discovering relevant visual concepts from a given scenario (body of text). Using the discovered visual concepts, I experimented with different methods to create a weakly supervised object detection method for these visual concepts using the Open Images dataset.

Semantic sketch parsing of sketches using Deep Hierarchical CNN
Prof. Venkatesh Babu, Indian Institute of Science, Bangalore
  • Model designed as a two level deep-network architecture. The first level contains shared layers common to all object categories. The second level contains a number of sub-networks.
  • Test sketch is classified by a super-category classifier and the resulting label is used to route the sketch to one of the super-category sub-networks.
  • Data-set with part-annotation information exists for natural photos but not for sketches. Therefore, for training, we use a sketchification process to make Pascal Parts data-set suitable for our purpose.
  • New data-set of part-annotated object sketches across 11 object categories created for evaluation of our model.
  • Paper published at ACM Multimedia (ACM MM) 2017. Code available here.


Online writer identification (Bachelor Thesis Project)
Prof. Suresh Sundaram, IIT Guwahati
  • Developed novel histogram based features to represent handwriting characteristics of a writer. A codebook was developed using sparse coding, modified Tf-Idf approach was used to create document descriptor and a one vs. all SVM model was used as classifier.
  • Improvement of over 10% obtained over state of the art results for the IAM On-Line Handwriting Database when using only online features at the text-line level.
  • Paper published with oral presentation at International Conference on Frontiers in Handwriting Recognition (ICFHR, oral) 2016. Code available here.

Approach to find vertices of the tree structure formed by particle showers (Summer Internship, 2015) 
Prof. Vincent Boudry, LLR, Ecole Polytechnique
  • Worked on reconstruction of objects in particle showers (in C ++, using PCA to make a linear approximation of particle showers), based on Monte Carlo simulated data for the International Large Detector(ILD). Th ILD will be one of the 2 detectors to be used in the International Linear Collider (ILC). In a detector, high energy particles decompose to form new particles (generally 2 or more). This process goes on for newer particles formed forming a tree-like structure. I designed and implemented an approach to find vertices of these trees using simulated raw data read from the detector (ILD).


Snake Robot with passive wheels and active joints (Design Project) 
Dr. Prithwijit Guha, IIT Guwahati
  • Designed and manufactured a wirelessly controlled snake robot with passive wheels and active joint modules in a group project.
  • Zigbee communication protocol was used for wireless control and Arduino Mega 2560 was used for control of servo motors.
See video of our snake robot below!


General Type-2 Fuzzy C Means and Android App Development (Summer Internship, 2014)
Guide - Prof. Frank Rhee, Hanyang University
  • Used MATLAB to implement general type-2 fuzzy C-means clustering algorithm (and interval type-2 fuzzy C-means clustering algorithm) using Euclidean and Mahalanobis distances.
  • Developed a system level Android input method editor (keyboard) for Korean language. The keyboard also provides word predictions as we type. [keyboard documentation]

Fabricated micro spur gear moulds (Summer Internship, 2013) [report (pdf)] [poster (pdf)]
Guide - Dr. Vishal Dhamgaye, X-ray optics Section
  • Intern at Indus 2, a Synchrotron Radiation Source at Raja Ramanna Center for Advanced Technology. Fabricated micro spur gear moulds of minimum feature length 55 µm using contact X-ray lithography and UV lithography. These micro moulds can be further electrodeposited using Nickel to obtain standalone metal gears.