Project Suggestions

Image Colorization

 

http://webee.technion.ac.il/labs/cgm/Computer-Graphics-Multimedia/Undergraduate-Projects/2011/AutomaticScenePainting/AutomaticScenePainting.png

Project Goal:

 

Develop an algorithm, which colors a gray scale image automatically using database of million images (from the Web). The basic approach is based on transferring a color from a source image to the grayscale image by matching different kinds of information between the images and then fusing the colors.

More Details: http://webee.technion.ac.il/labs/cgm/Computer-Graphics-Multimedia/Undergraduate-Projects/2011/AutomaticScenePainting/ProjectWeb/

Scene Completion

 

teaser.jpg

Project Goal:

 

Implement the following image completion algorithm. The algorithm patches up holes in images by finding similar image regions in the very large database that are not only seamless but also semantically valid. The algorithm is entirely data-driven, requiring no annotations or labelling by the user.

More Details: http://graphics.cs.cmu.edu/projects/scene-completion/

 

Image Blending

image equation

Project Goal:

 

Implement and test a poisson image blending.

More details about the approach : http://www.cs.brown.edu/courses/csci1950-g/asgn/proj2/

 

Automatic Photo Album

 

http://www.whitegadget.com/attachments/pc-wallpapers/73600d1314935671-album-photo-album-photo-pic.jpg

Project Goal:

Design and implement an algorithm that automatically chooses a few images out of hundreds of pictures from a trip or an event, which represent most of them.

 

 

Image Categorization with Thousands of Categories

 

 

Project Goal:

 

Run multi-class categorization on 1000 categories of ImageNet. The goal of the project is to compare one-against-all heuristic of SVM to Neural network classification. You should first get familiar with Caffe

 

1.      Download the 1000 categories from here

2.      Download pretrained CNN from and run the categorization using the CNN (here)

3.      Extract Deep Leaning Features as explained here.

4.      Train one-against-all classifiers using linear SVM on the extracted features (use LIBSVM)

5.      Run one-against-all classification using learned SVM models.

6.      Compare the results of the two systems.

 

 

Identifying the validity of the image in Identification

 

Project Goal:

 

Recent methods use face images for identification instead of passwords in mobile devices and other systems. It was shown that placing a photograph in front of the camera fools the system and makes it to believe that its a real person. The goal of this project is to develop a mechanism that can discriminate between a real person and an image. You can use video, audio, or any other input that can be placed in a mobile device.

 

Interactive Recognition

 

http://www.geeky-gadgets.com/wp-content/uploads/2010/12/sing-language-kinect-hack.jpg

 

 

Project Goal:

 

The goal of this project is interactive recognition of gestures using Kinect camera. You can get many ideas for projects that use the Kinect from the following sites: OpenKinect, or KinectHacks. Note: The camera will NOT be provided. You can choose this project only if you have your own Kinect camera.

Additional Projects in Collaboration with Robotics Lab

 

The projects are based on developing algorithms for the following setups and applications.

 

1. The relevant autonomous hardware:

A. Cars (lego or smaller race cars).

B. Helicopters (possibly that shoot arrows or water bubbles)

C. Mini Quadcopters

D. Large quadcopters (e.g. DJI phantom, for outdoor).

E. Fixed wing plane (outdoor).

 

2. Trackers (that automatically follow the robots or the people):

A. RGB on board

B. Kinect (3-D)

C. RGB on walls (stereo or mono)

 

3. Controllers (that tells the robot what do to) using:

A. EEG (brain)

B. Eye trackers

C. Gestures (e.g. MYO or Kinect)

D. IR or radio signals

E. Voice (speech recognition).

 

Applications include: War games based on e.g. face recognition, Swarms (operate dozens of robots), guiding robots, searching objects in the forest, delivering Pizza for windows in Eshkol building or in hospital etc...

The challenges are usually to develop fast and accurate algorithm for path planning and visual recognition.

One of the main theoretical tool that we will use will be coresets.

For additional information contact Dr. Dan Feldman