The computervision-hw2 from ml-lab

EN.601.661 – Computer Vision - Homework 2

To View This

This is a readme written in markdown. You can access here to view it online.

How to Run

# The results will be found in ./results/
python hw2.py

We also provide the pre-computed outputs in the directory ./results

Development Environment

Package	Version
Python	2.7.10
NumPy	1.13.3
OpenCV	3.3.0
matplotlib	1.5.3
SymPy	1.1.1

The Corner Detection Results

The corner detection result of bikes1

The corner detection result of bikes2

The corner detection result of bikes3

The corner detection result of graf1

The corner detection result of graf2

The corner detection result of graf3

The corner detection result of leuven1

The corner detection result of leuven2

The corner detection result of leuven3

The corner detection result of wall1

The corner detection result of wall2

The corner detection result of wall3

The Matching Results using the Sum-of-Squared-Distance

bikes1 and bikes2
The matchings between the two images look very good.

bikes1 and bikes3
The matchings between the two images look very good.

graf1 and graf2
The matchings between the two images look very good.

graf1 and graf3
The matchings between the two images are not very ideal. This is probably due to the big rotation which change the appearance.

leuven1 and leuven2
The matchings between the two images are not very ideal. This is probably due to the big lighting difference between the two images.

leuven1 and leuven3
The matchings between the two images are not very ideal. This is probably due to the big lighting difference between the two images.

wall1 and wall2
The matchings between the two images are not very ideal. This is probably due to the fact that the walls have repeating patterns and everywhere looks similar.

wall1 and wall3
The matchings between the two images are not very ideal. This is probably due to the fact that the walls have repeating patterns and everywhere looks similar.

The Matching Results using the Simple-SIFT

bikes1 and bikes2

The matchings between the two images look very good.

bikes1 and bikes3

The matchings between the two images look very good.

graf1 and graf2

The matchings between the two images look very good.

graf1 and graf3

The matchings between the two images are not very ideal. This is probably due to the big rotation which change the appearance.

leuven1 and leuven2

The matchings between the two images look very good.

leuven1 and leuven3

The matchings between the two images look very good.

wall1 and wall2

The matchings between the two images are sparse but accurate.

wall1 and wall3

The matchings between the two images are not very ideal. This is probably due to the big rotation that changes the distributions of the gradients.

RANSAC and Stitching using Affine and SSD

bikes1 and bikes2
The alignment between the two images looks very good.

bikes1 and bikes3
The alignment between the two images looks very good.

graf1 and graf2
The alignment between the two images looks very good.

graf1 and graf3
The alignment between the two images is not ideal. This is because the matching results are not ideal, which is due to the big rotation that changes the appearance.

leuven1 and leuven2
The alignment between the two images is not ideal. This is because the matching results are not ideal, which is due to the big lighting difference between the images.

leuven1 and leuven3
The alignment between the two images is not ideal. This is because the matching results are not ideal, which is due to the big lighting difference between the images.

wall1 and wall2
The alignment between the two images is not ideal. This is because the matching results are not ideal, which is due to the fact that the walls have repeating patterns.

wall1 and wall3
The alignment between the two images is not ideal. This is because the matching results are not ideal, which is due to the fact that the walls have repeating patterns. Besides, the rotation is also big.

RANSAC and Stitching using Perspective and SSD

bikes1 and bikes2
The alignment between the two images looks very good.

bikes1 and bikes3
The alignment between the two images looks very good.

graf1 and graf2
The alignment between the two images looks very good.

graf1 and graf3
The alignment between the two images is not ideal. This is because the matching results are not ideal, which is due to the big rotation that changes the appearance.

leuven1 and leuven2
The alignment between the two images is not ideal. This is because the matching results are not ideal, which is due to the big lighting difference between the images.

leuven1 and leuven3
The alignment between the two images is not ideal. This is because the matching results are not ideal, which is due to the big lighting difference between the images.

wall1 and wall2
The alignment between the two images is not ideal. This is because the matching results are not ideal, which is due to the fact that the walls have repeating patterns.

RANSAC and Stitching using Affine and SSIFT

bikes1 and bikes2

Using SSIFT produces similar but a little bit worse result. This is probably due to that the second image is blurred thus has less clear gradients.

bikes1 and bikes3

Using SSIFT produces similar but a little bit worse result. This is probably due to that the third image is blurred thus has less clear gradients.

graf1 and graf2

Using SSIFT produces similar result.

graf1 and graf3

Using SSIFT cannot produce good result, either.

leuven1 and leuven2

Using SSIFT produces result that is much better than using SSD. This is because the lighting changes do not change the distribution of gradients too much.

leuven1 and leuven3

Using SSIFT produces result that is much better than using SSD. This is because the lighting changes do not change the distribution of gradients too much.

wall1 and wall2

Using SSIFT produces result that is much better than using SSD. But I think this is due to the mechanism we use to determine the matchings between two points.