CS 180 Portfolio

We start with three original images: a left view, a center view, and a right view. These images have overlapping regions that we'll use to stitch them together.

Left Image

Center Image

Right Image

To stitch the images together, we need to find the homography between each pair of overlapping images. A homography is a 3x3 matrix that describes the transformation between two planes in a projective space.

The process of recovering homographies involves the following steps:

1. Find corresponding points. For this section, I found them manually
2. Set up a system of linear equations to find a matrix H that will transform the corresponding points from one image to another
3. Convert into matrix equation.
4. Use a least squares solver to find H, your desired matrix to transform the points from image 1 to image2.

Next, we project the left and right images onto the main (center) view. After calculating the Homography, we transform all of the pixels from the first image onto the second image using our Homography matrix that we calculated from the earlier step.

Left warped

Right warped

Finally, we combine all three images to create our final panoramic mosaic. This involves blending the overlapping regions to create a seamless transition between the images.

Left Image

Center Image

Right Image

Final Stitched Mosaic

The resulting mosaic combines all three original images into a single, wide-angle view, effectively creating a panorama that captures a broader scene than any single image could.

Now, we will follow the paper &quotMulti-Image Matching using multi scale oriented patches&quot to do autostitching instead of manual stitching. The image we will use is now this image.

Sproul Hall, Berkeley

Following the paper &quotMulti-Image Matching using Multi-Scale Oriented Patches&quot by Brown et al., we implement the following steps:

1. Harris Interest Point Detector: We implement a single-scale Harris corner detector to identify key points in our images.
   

   Sproul hall with Harris Points

2. Adaptive Non-Maximal Suppression (ANMS): We apply ANMS to select a well-distributed subset of the detected corners. The way we do this is to find the points that have both the highest definition (as calculated by our H function) and are both sparse by introducing the concept of a &quotsuppression radius&quot, which ensures sparsity between points.
   

   Sproul Hall with ANMS Points

3. Feature Descriptor Extraction and Matching: We extract feature descriptors from the selected points and match them between image pairs.
   

   feature matching

As we can see, there are some pretty good matching features, but there are also some features that do not match up well. In the below image, there is a more pronounced error in matching. This would result in a very bad translation, which is why I ended up ditching the below picture.

bad feature matching

Here are a few results! I ran into a few difficulties with my blending function, as well as picking better pictures. The way Ransac works is to randomly select 4 points and then calculate the homography, and if the distance between the original points and the transformed points are within a certain threshold, we call it an inlier. After a few times, we pick the homogrophy matrix with the most number of inliers.

Sproul Hall

Left Image

Center Image

Right Image

Manual Stitch

Auto Stitch

My Room

Left Image

Center Image

Right Image

Manual Stitch

Auto Stitch

My Couch

Left Image

Center Image

Right Image

Manual Stitch

Auto Stitch

By far the worst image was my couch -- I feel like there were not enough defined edges, having lots of detail at the edges.