HW #5 Projects

---

• Ahmad Faiz Abdull Rashid and Victor Ferrero
  Exploring Applications of Microsoft's Xbox Kinect: Using a Depth Map and Feature Points as a Password
  As more and more personal information is stored in the cloud the need for enhanced security is at an all-time high. With increased computing power, hackers can now figure out passwords much more easily than before. This coupled with the added pay-off that hacking can have has created a situation ripe for stealing our personal information. The days of passwords as strings of characters are numbered, and we must look to alternate methods of securing our computer systems. In this paper, we analyze the effectiveness of using a picture of a person as that person's password. Specifically, we explore the potential effectiveness of using a depth map along with twenty feature points in the human skeleton to create a pose based password.

• Raghu Balasubramanian and Prathmesh Prabhu
  StareZoom
  This work describes a technique for presenting a high resolution picture on a lower resolution screen while maintaining fine detail in a small area of attention. Instead of zooming a focus area to fill the screen, we enable retaining perspective by blending the zoomed area into a static background image. Further a program tracks a user's attention point and zooms into it without the need for any other input.
  Project web page.

• Matt Bollom and Sami Kuray
  Creating User-Driven Cinemagraphs
  Cinemagraphs are a relatively recent media whereby a video clip is transformed into a still photograph with an area of repeated animation. Originally created through manual manipulation with photo editing software, automated tools are now being developed to assist in the creation of cinemagraphs. However, a number of the tools available today are either entirely automatic or do not sufficiently constrain the area of animation. This project aims to create an automated tool to create photorealistic cinemagraphs while still allowing the user to assert creative control over the final product.
  Project web page.

• Aaron Bregger
  GPU Real-Time Video Abstraction
  I present a well researched method for generating non-photorealistic image abstractions that enhance the contrast of interesting image features. The image is heavily stylized using bilateral filtering and cartoon edges are added to emphasize areas of higher contrast. A soft luminance quantization technique is implemented to add a stylistic animated cartoon shading effect. I implement the automated abstraction of color video in real time using a portable GPU shading language. The method implemented achieves much faster than real time performance on modern graphics hardware.
  Video of results. Project report.

• Spencer Buyansky and Chris Lindner
  Qualitative Image Sorting via Relative Attributes
  A common way to describe images is with binary attributes; however this method is restrictive. You can say that one image is "urban" and another is "natural" but how do you distinguish which is "more urban" or "more natural" than another? What if the image contains both settings? Inspired by a paper by Devi Parikh and Kristen Grauman, we propose a way to rank images based on more qualitative attributes. This can be used to sort images in a collection. For our implementation we will use the example of "natural" vs. "urban" images, but if implemented properly, it is possible to qualitatively analyze different potential attributes. We will use various quantifiers and our own ranking algorithm to rank images based on what we calculate their setting or subject to be. Our goal is to have the best sort possible, and tweak it to the point where it can sort as well as a human could manually. It may be hard to define what makes an image urban for example, but we will attempt using techniques like line recognition and analyzing image content and general color schemes. We will make use the Flickr API to download a sample set of images to run though our program which is coded in MATLAB.
  Project web page.

• Jon Goetz and Kyle Keprios
  Occlusion Segmentation by Parallax Motion Analysis
  In order to accurately predict the a scene from a viewpoint between images taken from a plane, one must form a composite image of other viewpoints and realistically create a mask of the occlusions visible from that point. In order to facilitate this process we aimed to create an automated process that uses parallax motion to estimate comparative distances to objects in order to split the occlusions into a separate layer to be combined separately from the first.
  Project web page.

• Frank Greve and Nick Sefiddashti
  Super-Resolution Image Restoration using Texture Synthesis
  As technology for digital imaging advances every day, we are able to achieve increasingly higher resolution images. However, there are countless older low-quality pictures where it would prove useful to be able to restore these pictures using high-resolution samples (without losing the original structure and content of the image). The driving method of this algorithm uses a texture synthesis algorithm similar to that of Efros and Freeman's "Image Quilting for Texture Synthesis and Transfer" [1]. After synthesis, we can further improve the restoration by adding blending techniques to mask any signs of fabrication. This so-called "super-resolution image restoration" process might then be applied to images taken with old/cheap hardware, damaged images, or low-resolution video, in order to achieve resulting output of higher quality.

• Emma Jelley and Kevin Newman
  Picasso Blue Period Image Transformation
  The "BluePeriod" program, written in MATLAB, accepts a user specified source photo and processes the image data to create an output image characteristic of the works Pablo Picasso created during his "blue period." In addition to the image to be transformed a sample blue period image is chosen by the user and inputted to the program. This image determines the range of blues expressed in the final output image. The "BluePeriod" program is designed to run on images compositionally similar to the majority of works created by Picasso during this period, images portrait-like in nature with subjects featured in the foreground. An overall "blue period" aesthetic is achieved through three successive methods. The first emphasizes shading within subjects' facial features via manipulation of the image's luminosity. The second transfers color from the blue period sample image to the output image using the methods described in "Color Transfer between Images" by E. Reinhard, M. Ashikhmin, B. Gooch, and P. Shirley. The third employs tooning to create a painterly effect within the output image. In addition to these methods an optional highlighting method may be employed to brighten final images that appear excessively dark.
  Project web page.

• In Hei (Chris) Ku and Zhengyang Zhang
  Human Night Vision Simulation for Daytime Images
  In cinematography, taking daytime scenes and emulating night point of view is a practice researchers have studied for years. However, in photography, taking a daytime image and emulating night vision has never been in the spotlight. In this paper, we propose a few basic algorithms to take a daytime photo (with some restrictions) and create a human night vision (scotopic) of the same scene.

• Tim Lancina
  Improving Eigenface-based Facial Recognition with Depth Data
  In this paper I examine the potential benefits of including depth data along with traditional RGB images on recognition rates for the Eigenface-based facial recognition method. The traditional eigenface method using only RGB images is not very robust to changes in various conditions, such as orientation, lighting and occlusion. The goal of this project is to examine the effect on facial recognition rates under various conditions when incorporating depth information in addition to RGB images. Images and depth data for the offline 'training' phase of the eigenfaces approach are captured using an Xbox Kinect. Using the captured images and data, recognition rates for the traditional eigenface method are compared to the depth only and color-depth combination based eigenface methods.

• Jacob Laska
  Red Eye Detection and Removal
  A common problem in modern photography is the artifact of red eye in photos. Red eye is the result of flash reflecting off the subject(s) eye(s). This annoying problem can be dealt with by combining a few techniques. By analyzing the digital photograph for regions of red, a group of candidate areas is defined. From there, these red regions are examined, and it is determined if it occurs within an eye. Correction is implemented at the pixel level by altering the color channels.

• Sam Miles and Louis Schiff
  Stachion Statement
  This paper describes the application that uses facial recognition to alter human faces by adding facial hair. The goal is to alter the face so that it cannot be identified with current facial auto tagging software. This is accomplished in a three-step process. First, Viola-Jones facial recognition is used along with a customized feature detector to isolate the upper lip region. The mustache image is then resized, transformed to fit the face, and placed in the optimal position in the region. Finally, the mustache image is blurred and blended onto the face. The results are run through Picasa, which can detect people and automatically assigns similar faces to those people, to see if they can still be recognized.

• Harsh Newar and Alex Schwartz
  Face Anonymization in Images and Videos
  In the context of sharing video surveillance data, a significant threat to privacy is face recognition software, which can automatically identify known people, such as from a database of drivers license photos, and thereby track people regardless of suspicion. It is possible for anyone to trace your location right now, the only things he need is your current face picture and the access to the public surveillance video. This paper introduces an algorithm to protect the privacy of individuals in video surveillance data by de-identifying faces such that many facial characteristics remain but the face cannot be reliably recognized and further applying it to video.

  Many ad hoc attempts, such as covering eyes or randomly perturbing image pixels, fail to thwart face recognition because of the robustness of face recognition methods. This paper talks about a new privacy-enabling algorithm, called K-Same and another version called K-Same Select, that limits the ability of any Face Recognition Software to reliably recognize faces while maintaining facial details in the images. Face anonymization has expanded from standard blurring and pixelation to more advanced techniques to remove identifiable features. Utilizing the k-same select algorithm as described in "Preserving Privacy by De-Identifying Facial Images" (Newton et al.) applied to a database of faces, a standard filter will be developed that can be applied to a stream of images from a standard AVI video clip. Results are analyzed by applying facial recognition detection and personal checking from volunteers for ability to recognize and identify the original source image.
  Project web page.

• David Pipkorn and Preston Weisbrot
  Steganography -- The Hidden Message
  The need to hide messages and to secretly or efficiently pass information has been around since ancient times. Today digital media; including pictures, sound files, movies, etc.; provides a rich environment to hide more information than meets the eye of a casual observer by exploiting unused or redundant data bits. This information can be used for clandestine messages to secret agents but can just as easily be used as a compression method to send additional messages within a single file, thereby reducing the amount of data to be sent. For this project we explore two methods for encoding/storing a message within the RGB pixels of a cover image without visual distortion. For both methods a message is broken down to individual components, converted into 8-bit binary values, encrypted using a simple symmetric Exclusive OR (XOR) encryption key, and then encoded in the cover image by changing the least significant bit of the pixel value. The first method implemented sequentially stores the message starting with the top left pixel and then encodes the message from top to bottom and left to right. The second method implemented randomly encodes pixels across the entire image in an attempt to be less noticeable to analysis.
  Project web page.

• William Seale and Monica Thompson
  Testing, Tuning, and Applications of Fast Physics-based Fog Removal
  Physics-based fog removal is the method by which a standard atmospheric model of haze is used to try to estimate fog density in a given region of an image. The model first estimates the skylight level of the image by performing a region-based intensity search of the image. Atmospheric veil is determined by the intensity maximum of a given pixel's normalized RGB values and a haze map is created based on the varying levels of this intensity to estimate the varying degrees of fog. The haze map is then inverted and applied to the image to remove the "fog" and keep the surface albedo image behind the haze. The technique outlined in "Physics-based Fast Single Image Fog Removal" used standard models and applies approximations of traditional transforms to produce a fast estimate of these fog and albedo maps. In this paper we reproduce Yu, et al. techniques, and further tune their technique and reformulate the underlying mathematical estimation model to allow applications beyond simple fog removal.

• Rajan Shah and Nick Welker
  Coloring Book Generator
  Given a single image, the coloring book generator will create a coloring book style output image. Coloring book style images are pure black and white outlines of input images, so people can color in the objects in between the lines to produce creative works of art that demonstrate their own view of a photograph. The types of techniques that work the best depend on the given image. The generator outputs six different images with varying combinations of techniques, and ideally at least one of the output images will be a desirable one. Parents can take pictures of things that they may find significant and feed it to the coloring book generator in order to create a great activity for their children. This type of technique does not seem to be common; we could not a find an instance of a software program or feature that creates just a black and white outline from a photograph.
  Project web page.

• Changmin Shen and Ilya Yavnoshan
  Unblink
  One of the most common problems with group photos is having some people in the photo with their eyes closed. Since blinking is a quick action, taking photos in quick succession should allow us to obtain a photo with no member of the group blinking. Unblink aims to fix this by directly focusing on the eyes in every picture and getting rid of blinking. The central method used to create the output focuses on exploring differences in images to find if eyes have been blinking or not.
  Project web page.

• Zach Welch
  Imitating Artistic Styles using Image Analogies
  Image Analogy is a framework for image processing which uses a source pair of images as training data to \learn" an image transform and then applies this transform to a target image to create a new image. Image Analogy has been shown to have a wide array of applications. One interesting application is using works of art as training data, which can produce images that imitate the artistic style of the input artwork. This paper discusses the basic algorithm for Image Analogy, as well as a number of improvements and extensions that can be applied to produce higher quality images. This paper also looks at the different output images generated from a MATLAB implementation of Image Analogy when several key features are varied, as well as their effect on computation time.

• Zach Wernberg
  Sleep Tracking with Kinect
  Tracking motion has many practical applications in the real world. Many different techniques have been used for tracking motion while people sleep, but are often done using expensive cameras. In this paper I show a solution using an inexpensive Microsoft Kinect to achieve the results.

  ---