• If you are citizen of an European Union member nation, you may not use this service unless you are at least 16 years old.

  • Whenever you search in PBworks, Dokkio Sidebar (from the makers of PBworks) will run the same search in your Drive, Dropbox, OneDrive, Gmail, and Slack. Now you can find what you're looking for wherever it lives. Try Dokkio Sidebar for free.


Master's Projects (MCOGSc) ideas

This version was saved 10 years ago View current version     Page history
Saved by Jim Davies
on November 25, 2012 at 4:20:47 pm

Soon Carleton will be offering a master's in cognitive science (MCOGSc).

This page contains masters-sized project ideas.


Analogical transfer of motion between different 3D models

     This project, co-supervised with Dr. Ali Arya, will look at the case-based/analogical issues associated with transferring a motion-captured sequence (e.g., walking sadly) from one model (e.g., a person) to another (e.g., a dog). Other possibilities are transferring the dance style of one dancer to a different dance on another, or making any arbitrary 3D model walk similarly to any other walking 3D model given.


Machine Learning on Quanty Game Data to Improve Detectors

    This project is only possible after data has been collected from the Quanty Game, which, as of March 2010, is not yet live. The SOIL currently has implementations of spatial relation detectors (e.g., above/below, occlusion, close-to), based on computational, linguistic, and some psychological analyses. But they are not data-driven. This project involves making superior detectors based on data collected from human beings. It would involve using a portion of the Quanty Game data for training, and the rest for testing.


The next project ideas are based on the software Visuo, originally written by former lab member Jonathan Gagne. It takes in quantitative information (e.g., size, height) and outputs guesses for quantitative information for newly imagined things. It currently works with a database of tagged images. We have 50k images, many of the pixels of which are associated with labels (e.g., tree.)


Improving Visuo: Part-Whole Relationships (Kae Bagg's project)


Improving Visuo: In Contact

    Right now the distance between two labels in the 2D image is calculated as the distance between the average pixel location. This is not great, because they might overlap-- a smarter system would take into account the edges of the labeled area's outline. For example, how close are the closest points associated with the label? What things are adjacent to each other?


2D Image Choice Refinement: Occlusion

Right now SOILIE (SOIL Imagination Engine) grabs subsets of other images. For example, it might find an image with a house label in it, and pull out those pixels associated with that label. However, often the object is cut off by the side of the image. For example, a house or building might only be partially in view. This can be fixed by not using labels that have pixels up against the sides of the image. However, some things are almost always against the sides of the image, such as ocean, sky, road, horizon, etc. Find a workaround.

   Also, one image might be occluding another. A man sitting behind a desk will not be a complete image of a man. We can detect these cases by looking to see if the two objects share an edge. If so, the image must be ignored or the occluding object must come with it-- in which case we'd need to check to make sure that the new objects also coherently fits in the image we're constructing.  


Improving Visuo: Coherence In Search Results (Vertolli)

     Visuo uses the oracle of objects, which returns the top 10 labels that co-occur with some query label. For example, a "dog" might be associated with the labels "leash" and "fire hydrant." However, also in the top ten might be labels such as "sofa" and "dog bowl." But we know that a leash and a sofa are rarely going to be in the same image. This project's goal is to improve the search results so that the returned results are coherent-- that is, not only do they correlate with the query, but they correlate with each other to some degree. 


Improving Imagination Engine: Photo stitching

     Right now the imagination engine puts pixels from different images into a canvas, but there is a lot of blank space around the images. There are software techniques from computer graphics that "stitch" photos together so they look they all came from the same photo. This needs to be implemented for our imagination engine for a good demo.


Improving Visuo: Training Visuo (Breault)

     Visuo needs to be trained on all labels. This takes forever and needs to be babysat. Someone needs to set up scripts and make it happen.


Improving Visuo: Optimization (Gagne and Somers)

     Training takes forever, and the code can probably be optimized. 


Improving Visuo: Clean up LabelMe Database

     LabelMe is a publicly available database that we use to train Visuo, but it has a lot of junk. Set up a Amazon mechanical turk task that has people clean it up. Or find some other input that can replace it. 


Language to Knowledge Representation Game (AI, HCI)

     In this project you will work with Cycorp, Jim Davies, and Nancy Salay of Queen's University to make a game in which people turn knowledge representations in CYC-L into English, using "human computation."

Comments (0)

You don't have permission to comment on this page.