4th year MTech+PhD Dual Degree
Computer Science and Engineering
Indian Institute of Technology Bombay

Email: [firstname]m[at]cse.iitb.ac.in
Github: https://github.com/pmkalshetti

My thesis topic is Hand Tracking for Virtual Reality and my advisor is Prof. Parag Chaudhuri. The research involves Computer Graphics, Computer Vision, Optimization and Machine Learning, and heavily uses C++ (Eigen, Ceres-solver, Opencv, Open3d, pybind11) and Python (Tensorflow, Scipy).


  • 3rd place in ACM Student Research Competition at SIGGRAPH 2019, Los Angeles. Announcement
  • Awarded TCS Research Scholar Fellowship 2019.
  • Winner Qualcomm Innovation Fellowship India 2017. Results
  • 2nd place in Inter IIT Tech Meet 2014.


  • Kalshetti P, Chaudhuri P. "Unsupervised incremental learning for hand shape and pose estimation". In ACM SIGGRAPH 2019 Posters 2019 Jul 28 (pp. 1-2). https://doi.org/10.1145/3306214.3338553
  • Kalshetti P, Bundele M, Rahangdale P, Jangra D, Chattopadhyay C, Harit G, Elhence A. "An interactive medical image segmentation framework using iterative refinement". Computers in Biology and Medicine. 2017 Apr 1;83:22-33. https://doi.org/10.1016/j.compbiomed.2017.02.002 (Accorded Honors status!)


Some of my projects are listed below.

Point Cloud Sampling using Graph Signal Processing

This project uses graph signal processing to sample a point cloud such that application-dependent features are preserved. Initially a graph is constructed from the point cloud. A sampling distribution is then computed using this graph and the desired application. More precisely, the application decides the type of graph filter to be used for computing the distribution. Finally this distribution is used to sample the point cloud.

Code Link: https://github.com/pmkalshetti/fast_point_cloud_sampling

Fit Mesh to PointCloud

The algorithm is expressed as energy minimization. The energy is written as a sum of squares that is then optimized using Levenberg-Marquardt. For this project, the initialization is provided manually, however this can be provided by a discriminative model. The key novelty is to jointly optimize over both model parameters and correspondences between observed data points and the model surface.

Code Link: https://github.com/pmkalshetti/fit_mesh_to_pointcloud

Object Detection

Single Shot MultiBox architecture was used to solve this classification + regression problem and the variable dimensional output was handled using anchor boxes, resulting in real-time accurate detection.

Code Link: https://github.com/pmkalshetti/object_detection_old

Medical Image Segmentation Tool

In order to obtain the most suitable method for medical image segmentation, we propose MIST (Medical Image Segmentation Tool), a two stage algorithm. The first stage automatically generates a binary marker image of the region of interest using mathematical morphology. This marker serves as the mask image for the second stage which uses GrabCut to yield an efficient segmented result. The obtained result can be further refined by user interaction.

Paper Link: https://doi.org/10.1016/j.compbiomed.2017.02.002