Shaunak Srivastava

• Working in 3D vision using Generative Adversarial Networks (GANs) for realistic 3-D Face Synthesis.
• Involves using 3D Morphable Models (3DMMs), facial meshes, point-clouds and albedo-texture maps, UV position and texture mapping along with 3D rendering.
• Explored Auto-encoders, Variational Autoencoders (VAEs), GANs for non-linear neural generative modelling. This further explores effective Latent-space encoding, Feature Disentanglement and defining task specific Loss Functions for Supervised and Adversarial Training (e.g. Cross Entropy, GAN loss, WGAN-GP).
• Responsibilities also included preparation of Dataset/Data Loader, training & testing scripts along with Evaluation Metrics to measure performance.
• Also explored GPU Parallelism, optimizing runtime performance for our NVIDIA GPUs running CUDA with PyTorch. GPUs hosted on remote servers accessed via SSH.

Technologies used:

• Generative Adversarial Networks
• 3D Vision
• Differential Geometry
• 3D Face Modelling
• PyTorch
• 3D Rendering

• Researched on Statistical Process Control and its use in optimizing Six Sigma Processes.
• Analyzed manufacturing process data to find erroneous variations using statistical tools.
• Conducted statistical studies to find Process Capability (Cp), Process Capability Index (Cpk).

Technologies used:

• Minitab
• Statistical Process Control
• Probability & Statistics

• Created a Reinforcement Learning agent using NEAT (Neuro-Evolution of Augmenting Topologies) for environment exploration and collision avoidance. NEAT-python implements an evolutionary neural network to perform reinforcement learning.
• NEAT attempts to build an Artificial Neural Network (ANN) by adding and deleting neurons and modifying connections in a stochastic manner and evolving these networks/genomes as they reproduce through the generations.
• Experimented with different environments and reward functions to understand the effect of obstacle layout design on successful path planning and learning speed.
• Used ‘pygame’ library for creating the game environment.

Acheivements:

The agent was able to learn to avoid obstacles and also reach the goal using different strategies which exhibits its ability to explore the environment. The agent acheived this in different obstacle layouts.

Technologies used:

• Reinforcement Learning
• Evolutionary Neural Networks
• Collision Detection
• pygame
• Python

• Designed an algorithm for online Multi-Object Tracking which has been tested on the MOT Challenge benchmark and the KITTI dataset.
• Conducted a literature survey and study of various online tracking algorithms including SORT and DeepSORT.
• Explored CNN and color histogram-based feature descriptors for data associations.
• Worked with algorithms such as Kalman Filters, Hungarian Association Method, Linear Assignment, Feature Extraction and Track Management.
• Implemented using Python, NumPy and OpenCV.

Acheivements:

Achieved 77Hz real-time online tracking on the MOT16 benchmark with comparable accuracy (MOTA) performance to DeepSORT, illustrating improved computational efficiency.

Technologies used:

• Deep Learning
• Feature Extraction
• Kalman Filters
• OpenCV
• Numpy
• Python

• Trained a Face Recognition model implementing the Siamese Network to learn the use of triplet loss/contrastive loss for One-Shot Learning.
• Created a hand-gesture recognition tool for interactive gesture control using Google’s mediapipe for handdetection and classical techniques for tracking with OpenCV.
• Object detection using YOLO for object detection and classification on COCO dataset.
• Object detection and blurring using Haar Cascades for privacy protection using OpenCV.
• Object tracking using Lucas-Kanade Sparse Optical Flow.

Technologies used:

• Image Processing
• OpenCV
• Deep Learning
• Python

• The University Rover Challenge (URC) by the Mars Society has an autonomous driving task that requires the rover to autonomously navigate from the given GPS coordinates of the start and end point.
• We created a working simulation which demonstrates autonomous navigation of a vehicle on a path with obstacles.
• Used the ROS framework and Python along with Gazebo for the simulation environment.
• Used Sensor Fusion of multi modal sensor data, such as 3D depth data from RGB-D sensors, IMU data and GPS data for the simulation.
• Used algorithms like GMapping for SLAM and Extended Kalman Filters (EKF) and AMCL for localization.
• Path planning was implemented using move_base package.

Technologies used:

• SLAM
• ROS
• Sensor Fusion
• Path Planning
• EKF Localization
• Python