Segmentation Documentation


Recurrent Convolutional Net

Iain worked on the machine learning approach to segmentation, mostly with the census. They got decent results, and started on a paper.


Jesse was the main one who worked on the RANSAC segmenter. He got it to do the census fairly well, but we are still (8/2019) working on improving it. This was for the census. Daniel has been working on it this summer, getting the points it outputs to convert to ranges etc.

Ransac is an image-processing based system that works from a template of points and adjusts them to a given image.

An important feature of RANSAC is that when it messes up, it messes up bad (so you know it did). Previously it would just subtly put in wrong points, but this has been changed.

For more information, click here: Ransac Segmentation


Stanley worked on a specific type of segmentation called MaskRCNN. This focuses more on finding specific things in the image (examples are usually finding an animal), and classifying on them. Stanley used this for death records to find areas that specifically said the cause of death, and it is very, very good at that. project

We are working with to create recognition modules for French death records. The early stages of this are all segmentation jobs, so we are focusing on that right now.

Stanley’s MaskRCNN system did something similar, but unfortunately only gets the cause of death part. We need to get a whole bunch of information from the death records (Multiple fields such as name, date, DOB, POB, and many others). The RANSAC segmenter may work better, as it is made to segment many fields, but it has limitations itself.

One idea is that we would pass it all through RANSAC, then ones where it messes up (which shouldn’t be hard to tell, RANSAC freaks out when it messes up), we pass to a machine learning-based segmenter.

Brian Davis is a grad student who has been working on a somewhat similar project. He could be a helpful resource.

handwriting/segmentation.txt · Last modified: 2019/10/18 23:35 by
Back to top
CC Attribution-Share Alike 4.0 International = chi`s home Valid CSS Driven by DokuWiki do yourself a favour and use a real browser - get firefox!! Recent changes RSS feed Valid XHTML 1.0