Feature-based Image Stitching Benchmark~ FISB

---

Report | Demo-video | Dataset | Presentation

Introduction

The repo presents feature-based image stitching algorithms for creating panoramic images from multiple input images. The proposed method involves feature extraction, matching, and blending stages. The Google Landmarks database and a custom dataset were used to evaluate the effectiveness of the proposed pipeline, which was measured using objective and subjective metrics, including accuracy, speed, and visual quality. The experimental results show that the proposed pipeline can effectively stitch images and produce seamless panoramas. The pipeline is scalable and can be used in various applications, such as surveillance, virtual reality, and cartography.

---

Demo video

Execution

Installing requirements with conda

Run the following command in the master root:

conda install --file requirements.txt

Custom Pipeline execution

The following command executes main.py for a custom analysis of algorithms.

Function Parameters:

• Select the Feature Matching Algorithm
• Choose the Feature matching algorithm
• Specify the Image Blending tecnhique in main.py for the usecase
• n: number of images, alpha: maximum feature descriptors (1<=alpha<=100)

python main.py

Automated FISB Pipeline Execution

Executing autoMain.py instead of main.py:

python FISB-Pipeline/autoMain.py {Feature Descriptor} {Matching Algorithm} {n} {alpha} fisb_dataset/sub/{}/ >> output/logs/log{}.txt

Executing script for the fisb dataset

python script.py

---

Dataset description

• Find dataset here
• Dataset has two folders: sub and super
• super has 49 natural and digital scenes
• sub has 49 folders containing sub-images of the scenes in super
• the sub-images are of varrying view-points, perspective, device, camera orientation and illumination

Note: Results of our implementation from using SIFT+BFMatcher+Seamless blending (best result parameters) have been cached here

scene_9.png sub-images of scene 9

---

Comparisons

Lets compare different blending results on scene_10

Original Image

Sub Images of Scene 10

• Some blending techniques employ different masking techniques. To compare on equal footing, the final stitched image is cropped and presented as below

Alpha Blending

Gaussian Blending

Multiband Blending

Seamless blending

Result given by OpenCV

• For more comprehensive analysis and comparison, refer to report

---

Contributors

Abhishek Rajora [email protected] ; Github: brillard1

Abu Shahid [email protected] ; Github: ceyxasm