UConn Voter Center - Voting Bubbles with Marginal Marks

1. Pre-Print Dataset

This set of bubble images are the original digital images of empty and filled bubbles. This category is further split into the following sub-categories:

• OnlyBubbles: This dataset contains images of empty and filled bubbles (class 0/1) from the original dataset.
• OnlySynthetic: This dataset contains synthetically-generated marginal marks (penrest, check, cross, straight scribble, random scribble) superimposed on top of blank bubbles.
• Combined: This dataset is a combination of OnlyBubbles and OnlySynthetic.

  Note that for OnlySynthetic and Combined, we performed manual handlabeling of the first 100 synthetic marks of each class sorted from lowest to highest pixel intensity for pen rest, cross, and check. It was sorted from highest to lowest for Straight and Random Scribbles. Some synthetic marks were reassigned to the opposite class. This only occurs for Cross and Checks, reassigning from class 1 (filled) to class 0 (empty) as they were too light/visually imperceptible/ambiguous to be classified as a filled bubble.

  
  
  Each dataset is split into train/validation sets. See the table below for dataset split specifications.
   

  Train/Val	Data Combination	Grayscale/RGB	Data Size
  Train	OnlyBubbles	Grayscale	28,589
  Train	OnlyBubbles	RGB	28,589
  Train	OnlySynthetic	Grayscale	14,090
  Train	OnlySynthetic	RGB	14,090
  Train	Combined	Grayscale	42,679
  Train	Combined	RGB	42,679
  Validation	OnlyBubbles	Grayscale	7,137
  Validation	OnlyBubbles	RGB	7,137
  Validation	OnlySynthetic	Grayscale	3,515
  Validation	OnlySynthetic	RGB	3,515
  Validation	Combined	Grayscale	10,652
  Validation	Combined	RGB	10,652

  
  

  2. Post-Print Dataset

  
  
  This set of bubble images have been printed once and scanned again. We use the following equipment for printing/scanning:
  • Printer:
    • Grayscale: HP LaserJet Pro MFP 3101fdw
    • RGB: Brother HL-L8360CDW
  • Scanner: Ricoh fujitsu fi-7160
  
  
  We produce the following post-print datasets. Note that some do not contain the same amount of images as their corresponding pre-print versions, they are abbreviated versions. We anticipate producing a larger and complete range of post-print data in the future.
   

  Train/Val	Data Combination	Grayscale/RGB	Data Size
  Train	OnlyBubbles	Grayscale	2,000
  Train	Combined	Grayscale	42,679
  Validation	OnlyBubbles	Grayscale	2,000
  Validation	OnlySynthetic	Grayscale	3,515
  Validation	Combined	Grayscale	10,652
  Validation	Combined	RGB	10,652

  
  

  Examples of Dataset

  See the linked Github repo (https://github.com/VoterCenter/Busting-the-Ballot/blob/main/dataset/read_dataset_example.py) for how to read in these datasets.
  
  
  Here we include an example of marks from each dataset class from `Combined`. Note that OnlyBubbles + OnlySynthetic = Combined. We also include the original labels for each class and their binary mapping which we use in our dataset.
   

  Mark Type	Binary Label	Original Label
  Empty	0	0
  Filled (original)	1	1
  Penrest	0	2
  Check	1	3
  Cross	1	4
  Straight scribble	1	5
  Random scribble	1	6