eruka-cmu-housing/historical-appraisals-ocr-ml

Digitized Historical Property Valuations and Building Feature Data

A dataset of historical property valuations and contemporary parcel/building feature data created by the process described in this ACM COMPASS paper: TODO add link when published

Data description

hamilton_county

Data for properties in Hamilton County, Ohio (primarily Cincinnati). All the data sources below are linked using the land parcel identifier parcelid.

• property_cards/: folder with tar balls of scanned images of the historical property valuation cards that we digitize, with each image name indicating its parcelid. Sourced from the Hamilton County Asessor here.

• hand_annotations_single_cell_hamilton.csv: human annotations for a random sample of 12,269 cards. Specifically, these capture the earliest "building" value from the historical cards, which corresponds to the year 1933 (hence single cell, rather than every row/assessment). The "handwritten" column indicates if the value on the card was handwritten or printed.

• ocr_predictions_single_cell_hamilton.csv: contains predictions from our OCR pipeline, along with model confidence score, of the 1933 "building" value. Confidence score ranges from -0.630936 to -0.030582 (higher is more confident). Predictions are made for ~50k property cards for which image segmentation to find the 1933 building value cell in the full card was successful.

• building_features_hamilton.csv: data from 2023 about each parcel, and detailed information about the building(s) on that parcel (e.g., home size, characteristics, etc.), sourced from the Hamilton County Auditor's property assessments here.

  • Data dictionary describing each feature and how it was processed is in this GSheet
  • SQL code that takes the raw data and transforms it into this cleaned version is in this folder of our repo.

franklin_county Data for properties in Franklin County, Ohio (primarily Columbus). All the data sources below are linked using the land parcel identifier parcelid.

• property_cards/: folder with tar ball of scanned images of the historical property valuation cards that we digitize, with each image name indicating its parcelid. Sourced from the Franklin County Assessor here.
• hand_annotations_single_cell_hamilton.csv: human annotations for a random sample of 672 cards. Specifically, these capture the earliest "building" value from the historical cards, which corresponds to either the year 1920 or 1931 (hence single cell, rather than every row/assessment).
• building_features_hamilton.csv: data from 2023 about each parcel, and detailed information about the building(s) on that parcel (e.g., home size, characteristics, etc.), sourced from the Franklin County Auditor's property assessments at this link
  • Data dictionary describing each feature and how it was processed is in this GSheet, in the Franklin County tab
  • SQL code that takes the raw data and transforms it into this cleaned version is in this file of our repo.

ocr_data_comprehensive_method

This dataset also includes a Comprehensive Format OCR Dataset, which was used in the comprehensive digitization technique - for training and evaluation of finetuned TrOCR model.

• images/: Folder of cropped cell images, named in the format parcelid_cellnumber_label (e.g., parcelid_3_555.jpg), where:
  • parcelid is the land parcel identifier.
  • cellnumber refers to a manual cell number corresponding to the location of a specific field on the full property card. You may refer this GSheet, first subsheet 'template' contains the pixel coordinates of each cell position so that each cell can be extracted from a property card, after alignment with the empty template. Second subsheet 'numbered' contains the mapping of a cell number to its corresponding pixel coordinates. This is where we get cellnumber from.
  • label is the human-annotated value for that field.
• labels.csv A labels file that contains two columns:
  • parcelid (which is the cropped image filename present in images/)
  • value (the human-labeled field value)

Relevant links

GitHub repo: link Paper: TODO link

Dataset Card Contact

Mihir Bhaskar (https://mihirbhaskar.github.io/), Jun Tao Luo (https://github.com/JunTaoLuo)