Description
In this project, you will train a model to detect barrels in images and find the relative world coordinates
of the barrel from images. To do this, you will need to implement algorithms that learn the color model,
segment the target color, and finally localize the target object from images. Given a set of training images,
you will hand-label some training examples. From these training examples, you will build a color classifier
and red barrel detector. You will then use your algorithms to mark the center of the detected barrel and
display the distance to the barrel on new test images.
Training Data Download: Now available at https://cornell.box.com/v/ECE5242Proj1-train
Test Data Release: 2/10/2022 9:25am https://cornell.box.com/v/ECE5242Proj1-test
Upload to Canvas:
1. Code (due 2/10/2022 9:25am (15 mins before class starts), [your netid] project1.zip )
• Do not include image data.
• Include your code
• Include a README file with clear instructions for applying your code to test images
• Include trained models that can be applied to the test images
2. Write-up (due 2/14/2022 11:59pm, [your netid] project1.pdf )
• Write the summary of your approach, result, and discussion.
• Make sure your writeup includes your result for each test image:
(a) the bounding box of the barrel,
(b) the center coordinate of the barrel,
(c) the distance estimate to the barrel for each test image.
Instructions and Tips:
1. Hand-label appropriate regions in the training images with discrete color labels. For this project,
we will be especially interested in regions containing the red barrel. If you are more ambitious, you
could try to implement more automated ways of labeling images – perhaps by an unsupervised image
segmentation, or an adaptive region flooding algorithm.
Tips: You can (but dont have to) use RGB color space. There are useful built-in functions if you
want to use a different color space. Lighting invariance will be an issue, so you should think carefully
about the best color space to use, and perhaps some low-level adaptation on the image.
2. Use a learning algorithm to partition the color space into appropriate class color regions. You should
first implement the approach discussed in class (you should implement it yourself, you should not
use sklearn.mixture.GaussianMixture), but you can also try other machine learning approaches, i.e.
decision trees, support vector machines, etc. You need to make your algorithm so that it is able to
robustly generalize to new images. To do this, you should use cross validation (hold out some of the
training images to test this). This will allow you to compare different parameters for the probabilistic
models and different color space representations.
1
3. Once the red barrel regions are identified, you can use shape statistics and other higher-level features
to decide where the barrel is located in the images. The filename gives [the number of meters away
fromthe barrel].[no meaning].png. If there are two barrels, the two distances are separated by an
underscore. Use your algorithms to label and identify the coordinates of the centroid of the barrel in
a new test image. You should also compute an estimate of the distance to the barrel. You can use
the fact that the barrel has dimensions 16 1/4” Diameter and 22 7/8” height if you need to. Youll
be expected to quickly be able to classify and display your results on a new set of test images. Write
a test script to display your result using something like:
import cv2, os
folder = “Test\_Set”
for filename in os.listdir(folder):
print(filename) # record the filename, since listdir is unordered
# read one test image
img = cv2.imread(os.path.join(folder,filename))
# Your computations here!
x, y, d = myAlgorithm(img)
# Display results:
# (1) Segmented image
# (2) Barrel bounding box
# (3) Distance of barrel
# You may also want to plot and display other diagnostic information
cv2.waitKey(0)
cv2.destroyAllWindows()
4. During your (optional) class presentation, you will be asked to run your code on the test set images
which will be released online prior to the presentations.
5. Your write-up upload should include your test results in the following manner:
• You have to include images of test set with centroid of barrel in your report!
• Centroid points and distances of test set. For example:
ImageNo = [01], CentroidX = 662.0208, CentroidY = 506.7571, Distance = 5.441
ImageNo = [02], CentroidX = 507.0605, CentroidY = 826.7004, Distance = 2.384
… etc.
6. You may use some useful python functions for this project:
• hand-labeling: roipoly: https://github.com/jdoepfert/roipoly.py
• color conversion: cvtColor: https://docs.opencv.org/3.0-beta/doc/py_tutorials/py_tutorials.html
• region analysis: regionprops: https://scikit-image.org/docs/dev/api/skimage.measure.html
However, please do not use any built-in function that would be the core part of this project. If you
are not sure, then ask a TA. Examples of allowed code:
import cv2
img2 = cv2.cvtColor(img, cv2.COLOR_BGR2YCR_CB)
from skimage import data, util
from skimage.measure import label, regionprops
img = util.img_as_ubyte(data.coins()) > 110
2
label_img = label(img, connectivity=img.ndim)
props = skimage.measure.regionprops(label_img)
import numpy as np
np.cov # any base numpy function (not scipy or sklearn) is OK.
Tips for the Writeup:
1. Format. You can any format for the report, but please try to keep to the outline below:
• An introduction
• A problem statement
• A description of your approach
• A discussion of your results
• references (If needed)
Please don’t include your code or screenshots of the code in the write-up report. If you want to
describe specifics about your algorithm, please use pseudocode.
2. Description for the approach in detail:
• how to decide color space
• how you picked your threshold(s)
• how to decide the number of GMM
• plot of a learning curve (loglikelihood plot)
• your model for distance to barrel
• etc
3