HW2

Due: 4/19 11:00am EEE Dropbox

Color-based object detection

In this assignment, you will make a pixel-based classifier for different body parts of a moving person. You will use the "color" images from the project video set. The classifier, constructed through Baye's rule, will classify a pixel "x" as foreground if P(x|foregound)P(foreground) > P(x|background)P(background). You will explore different likelihood models for P(x|foreground) and P(x|background) using multivariate Gaussians and histogram models. You will use these models to detect the torso, hair, and legs of the person moving in the video.

Overview: You will be given skeleton code here. The two high level scripts "hw2a.m" and "hw2b.m" contains wrapper scripts for color-based pixel detection, using Guassian vs histogram color models. Your job is to fill in all the referenced functions to allow the given wrapper scripts to execute correctly. You should not need to modify the wrapper scripts beyong pointing to your directory of images.

User interaction: The wrapper scripts require a user to mask out an object in a given frame, using matlab's "roipoly" function. I have included a mask for the torso of the figure in zipfile, saved as a matlab matfile. There are three sections to this assignment.

Helper functions

  1. vectorize.m [10 pts]

    This is a helper function that flattens an image into an array of RGB vectors. This will be useful for passing pixels into the fitting functions below that learn color models.

  2. fitPriors.m [10 pts]

    This function will estimate the maximum likelihood estimate (MLE) of the prior probability of a foreground and background labels given a a vector of binary input labels.

Gaussian color models

  1. logGaussian.m [20 pts]

    This function will evaluate the log probability of seeing a collection of pixels under a given Gaussian model (eg, a mean RGB vector and 3x3 covariance matrix). This will be a helper function called by "classifyGaussian.m".

  2. fitGaussian.m [20 pts]

    This function will learn a Gaussian color model (eg, a mean RGB vector and a 3x3 covariance matrix) given a collection of vectorized pixels.

  3. classifyGaussian.m [20 pts]

    This function will classify an image using a foreground and background Gaussian color model, and prior models for foreground and background labels.

Histogram color models

  1. quantizeIm.m [20 pts]

    This function will quantize an image, so that each pixel is labeled with the bin it falls into. If quantizing into 8 bins per color channel, each pixel will be labeled with a number between 1 and 512. This helper function will be called by the bottom two functions.

  2. fitHistogram.m [20 pts]

    This function will learn a discrete probability model given a collection of pixels by counting the fraction of times a given discrete value is observed. By default, this will assign zero probability to a discrete value that is not observed. You can fix this by adding a small value to the count of every value.

  3. classifyHistogram.m [20 pts]

    This function will classify an image using a foreground and background histogram color model, and prior models for foreground and background labels.

What to hand in: Hand in all the completed functions above, complete with comments. Also hand in a pdf with the following figures. To create the figures, select four frames from the video. You pdf should include figures of these images and their associated pixel classification results.

  1. Learn color models for the hair, torso, and leg of the person in the video from the first frame. Show results for hair, torso, and leg detection on three remaining images from the video. You will need to use "roipoly.m" to generate label masks for the body parts, as shown in the skeleton code. Show results using a Gaussian color model, a 8-bins-per-channel color histogram model, and a 16-bins-per-channel color histogram model. [10 pts]

  2. Use the first two images for training, and the last two for testing. Again, use "roipoly.m" to generate the additional labels. Show the above images again, and explain why the results look better. [10 pts]

  3. [Extra-credit] The binary classification masks will look noisy, with lots of "speckles". Use Matlab's morphology functions ("imopen.m" and "imclose.m") to clean up the classifications by exploiting the fact that nearby pixels should have similar labels. [10 pts]

Hints