SFM Learner

Currently estimating depth and pose (or ego-motion) from a sequence of images using unsupervised learning. Using multiple image sequences from KITTI dataset and learning rotation and translation from the sequence.

Learn more

DepthSegNet - Monocular Depth estimation and Semantic Segmentation

This project addresses the two most important task in computer vision of depth estimation and semantic segmentation. We have considered the idea of integrating both the task together in the same framework as they may benefit each other in improving the accuracy.

Learn more

AutoPano - Deep Homography Net, Supervised and Unsupervised

Developed a pipeline to implement a deep CNN to learn homography using TensorFlow with a custom-built dataset based on MS COCO thus generating a panorama using image stitching. Also, Performed the same task with an unsupervised homography net using TensorDLT and Spatial Transformer Network.

Learn more

Auto Calib

Estimating camera parameters like focal length, distortion coefficients and principal point automatically

Learn more

Face Swap (Snapchat Filter)

Created an end to end pipeline to swap faces in a video like Snapchat’s face swap filter. Also, Implemented a joint 3D face reconstruction and dense alignment network with position map regression. Also, implemented a classical computer vision pipeline using Delaunay Triangulation, Thin Plate Splines.

Learn more

Object segmentation

Programed a PR2 to pick and place objects through the video feed of an RGB-D camera. Used PCL, RANSAC based plane segmentation and DB Scan Clustering for object segmentation. Implemented RBF SVM, for object classification, using color and surface normal histograms.

Learn more

Vehicle detection for autonomous driving

Implemented a CNN for a car detection pipeline using YOLO (You Only Look Once) algorithm in TensorFlow and Keras.Used Non-maximum suppression and other techniques for improving the accuracy of detection

Learn more

Image Classification on Fashion-MNIST, CIFAR 10 dataset

Implemented maximum likelihood estimation with Gaussian assumption followed by Bayes rule for classification. Applied PCA, LDA for dimensionality reduction, and then classified the images using KNN, and SVM. Implemented LeNet, VGGNet and ResNet, ResNext architectures using TensorFlow as well.

Learn more

Neural Style Transfer

Implemented a Neural Style Transfer Algorithm and generated novel artistic images using that algorithm.

Learn more

ResNet

Implemented the basic building blocks of ResNets and put together these building blocks to implement and train a state-of-the-art neural network for image classification.

Learn more