thabsheerjm / my_autopano

Autopano - Stirching multiple images to form a seamless panorama

My AutoPano - Image Stitching

In this project we are learning to stitch two or more images in order to create a seamless panorama by finding the homohraphy between the two images.

Phase 1 : Classical Computer Vision approach

The claasical method involves several steps, it starts with corner detection. Then Adaptive Non-maximal Suppression (ANMS) is applied to ensure an even distribution of conrners. Feature descriptors are created by encoding information at each feature point into a vector. The next step is feature matching, where feature points between images are matched. RANSAC is used for outlier rejection and to estimate a robust homography. Finally, the images are blended to produce the panorama.The deatiled steps are explained here.

How to Run:

Place the image pairs in the dir:"../Data/Train/Set1/" and run the following command:

 python3 Wrapper.py

The classical pipeline is given below:

Corner Detection

Non-Maximal Suppression

Feature Matching

Outlier Rejection

Parorama: Warping, Blending and Stitching

Phase 2: Deep Learning

The deep learning model effectively combines corner detection, ANMS, feature extraction, feature matching, RANSAC and estimate homography all into one robust generalizable network. The complete methodolgy is given here.

How to Run:

Run the following to generate the dataset

python3 Datagen.py

Divide train and validation. and run:

python3 network.py
python3 Train.py
python3 Test.py

The network pipeline is given below:

We followed a supervised model and obtained the follwing results:

The following picture shows the input and output patch from the trained network.

The training loss over number of epochs is given below:

Acknowledgement

This project was part of RBE549- Computer Vision (Fall 22) at Worcester Polytechic Institute[1].
Team members :Thabsheer Machingal and Krishna Madhurkar

References

[1] RBE549-Project1