My AutoPano

The purpose of this project is to stitch two or more images in order to create one seamless panorama image by finding the Homography between the two images. The project is divided into two phases,

1. Phase 1: Classical approach of local feature matching
2. Phase 2: Deep Learning approach(Homography Net - supervised and unsupervised) to estimate the homography.

Team Members

• Rutwik Kulkarni ([email protected])
• Ankit Mittal ([email protected])

**Phase 1 - Using Classical CV **

Implemented traditional CV pipeline combines algorithms of corner detection, ANMS, feature extraction, feature matching, RANSAC, homography estimation and blending.

Results

Corner Detection and Non-Maximal Suppression

Feature Matching

Outlier Rejection using RANSAC

Warping, Blending and Stitching

Usage Guidelines

1. Open directory Phase1/Code and run the following command with the Data location as command line argument: -

   python3 Wrapper.py 
   

2. Results folder contains stitched images of all Train Sets and Test Sets.

Phase 2 - using Deep Learning

In Deep learning, used Homography Net (both supervised and unsupervised) to estimate the homography.

DataSet Generation

To generate dataset, run the following command in Phase2/Code/supervised: - python3 make_dataset.py

Original Patch

Patch after applying perturbations

Supervised Homography

Result

Training Loss

Validation Loss

Stitched Image.

1. To train the network, run: -

   python3 Train.py
   

2. To test the model on test set, run: -

   python3 Test.py
   

Unsupervised Homography

1. To train the network, run: -

   python3 Train.py
   

2. To test the model on trainset , run: -

   python3 Test.py
   

Result

Training Loss

Validation Loss

Output on the Patch (Image, Ground Truth, Prediction)