CS 194-26 Project 6 [acc id: aez]

Overview

The main idea was to shift all the images to align to the central image of position images[8, 8] out of a 17x17 grid of images. To vary the depth, I scaled the shifts appropriately
- The basic shifts were computed with the image coordinates (information embedded within the images of the Stanford Lightfield Archive)
- The scaled shifts were computed by scaling the basic shifts. I found that the scales of range [-0.4, 0.4] worked well
- Mathematically: shift = scale * (coords[8, 8] - coords[x, y])

Image
Scale	-0.4	0.04	0.4

A small aperature gives the image little depth of field, meaning that every object is clear. A large aperature gives the image depth of field, aka “focus/bokeh”.
- Small aperature is achieved by taking the average small number of images
- Large aperature is achieved by taking the average of most/all images
To simulate the effect of adjusting the aperature, I varied the number of images to be averaged as a function of the distance from the central point images[8,8] out of the 17x17 grid images.
- The distances to center based on range [0, 100] adequately captured the different aperatures

Image
Distance from center	10	50	100 (all images)

Probably the power/concepts of parallax in varying depth! Initially I thought that to vary parallax to refocus the image, you would have to vary which point all the images should align to. I later found that the scale is what affects the refocusing.
In that vein, I wonder if there would be any effects on varying where all the images are centered on. For this project, I picked images[8,8] and tried other positions but there were no noticeable changes. I am curious to see if this is true for other images.

Implemented by taking a simple squared difference between neighboring pixels. Another possible energy function would be a gradient function similar to that in Project 2.
This function is used to generate the energy matrix.

A cumulative energy path marix is generated from the energy matrix from the top to bottom to find vertical seams.
With the cumulative energy path matrix, the vertical seam is found simply by retracing back from the bottom to the top of the cumulative energy matrix by finding the minimum path.

I did the implementation to find and remove vertical seams, and simply transposed the input to do horizontal seam removal.

The image only had one main feature - what I expected the algorithm to do would be to trim the background. Instead, my bottle ended up being carved
- This could be due to the relative uniformity in color of the bottle as compared to the surrounding surface which captures shadows of light

Original	Carved

This is a smaller failure cases. But I noted that distortions in certain faces can be seen when more than 150 pixels were shaved off from the original image

Original	Carved (150px)	Carved (300px)

I implemented optimization by creating and offline and online function
- Offline: precomputes all the seams and saves them in an array
- Online: based on the target resize

I implemented seam insertion simply by identifying seams for removal as per in seam carving, and then inserted a seam at those target locations based on the average of the neighboring seams.

Original	Seam Insertion

When implementing seam insertion, I learnt that if you iteratively inserted seams based on the lowest energy seam, seams would always be inserted at the same location.
I followed the direction of the paper and instead inserted seams at positions based on seams identified for removal.

Lightfield Camera: Photos were taken from Stanford Lightfield Archive
Seam Carving: All photos are taken by me, except for the following taken from Google images
- Celebrity photograph (seen in part 3)
- Memorial glade (seen in both part 2 and part 4)