CS223B Homework 1 Results

1
CS223BHomework 1 Results
2
Considered 2 Metrics

• Raw score
• Number of pixels in error
• Weighted score
• Car pixels weighted more heavily than non-car
  pixels
• Range from 50-100
• Formula 40 ( of correct car pixels) 30
  (1.0 - of false positive pixels) 20 ( of
  correct non-car pixels) 10 (1.0 - of false
  negative pixels)

3
(No Transcript)
4
(No Transcript)
5
Best Solutions

• Eric Park, Brian Tran, Joakim Arfvidsson
• 3354 error pixels / score 84.3
• Fraser Cameron, Peter Kimball, Mike Vitus
• 3447 error pixels / score 77.2
• Simon Berring, Anya Petrovskaya, Daniel Tarlow
• 4337 error pixels / score 86.7
• Antoine el Daher
• 4518 error pixels / score 87.2

6
Eric Park, Brian Tran, Joakim Arfvidsson

• Road detection
• sample road color from just in front of car
• flood-fill the road using the sampled color
• use the RANSAC to find the edges of the road
• blur and threshold image
• Car edges detection
• Canny
• normalize edges
• extract horizontal and vertical edges from this
  result
• apply pattern matching
• Use perspective to dismiss false positives

7
Eric Park, Brian Tran, Joakim Arfvidsson
8
Eric Park, Brian Tran, Joakim Arfvidsson
9
Eric Park, Brian Tran, Joakim Arfvidsson
10
Fraser Cameron, Peter Kimball, Mike Vitus

• Road finder
• Prewitt edge convolution and a Hough Transform
• Tail light finder
• based on color
• Shadow finder
• looks for dark horizontal edges
• Box finder
• uses data from the above to generate bounding box
• Pixel classifier
• corner finding -gt convex hull to trace car edges

11
Fraser Cameron, Peter Kimball, Mike Vitus
Road Finder
Taillight Finder
12
Fraser Cameron, Peter Kimball, Mike Vitus
Shadow Finder
Box Finder
Pixel Classifier
13
Simon Berring, Anya Petrovskaya, Daniel Tarlow

• Ran four classifiers and combined the results
  using a naive Bayes model
• boosted Haar classifier detector
• color segmentation
• corner finding
• road finding

14
Simon Berring, Anya Petrovskaya, Daniel Tarlow
NaïveBayesModel

Haar Detector
Color Segmentation
CornerFinding
15
Antoine el Daher

• Trained several different boosted Haar
  classifiers
• 2 rear detectors
• 1 "far away car" detector
• 1 side cars" detector
• 1 "tail light" detector
• Ran a consistency checking phase
• Make sure car is in road region at a plausible
  depth, eliminate double detections
• Ran a refinement phase
• Tighten bounding box around car using "cube"
  model of car

16
Antoine El Daher
17
Antoine El Daher
Taillight Mask
Road Detector
End Result