Pipeline Overview - byuawsfhtl/RLL_computer_vision GitHub Wiki
BYU Handwriting recognition model Video
Train a Segmentation Model
We use Detectron2 to learn from the training data we create to be able to predict where the labeled things are in new images. We typically use Detectron2 on Google Colab, and move the trained model onto the BYU supercomputer.
Make Training Data
We use a program called Labelme that allows us to label images into different sections. It saves the coordinates of the different labeled elements in the images. This often gets outsourced to the Ghana team. However, it is important to know the basics as we occasionally need to make corrections. We will also sometimes do this ourselves due to time constraints.
Running Segmentation
We typically use environments running Detectron2 on the BYU supercomputer to segment images before we run HWR on them.
Handwriting Recognition (HWR) / Optical Character Recognition (OCR)
We then use Machine Learning to do HWR of the segments we get from Detectron2 if they are handwritten, or OCR if the record is printed. A portion of this is done by the Denmark team. However, we often take part in creating training data by indexing the data that will be used to train the model.
How to access the remote desktop / lab computers
See the wiki article for how to get that connection setup.
Some training videos
These videos are from past instructional zoom meetings. They may be helpful in figuring out how different parts of the lab work. You can also find various resources for coding help in the Supplementary Coding Resources page in this wiki.
https://byu.zoom.us/rec/share/6ku7w8nkh0tsxNQCq2aZHft5s2vTOoEhXSZVFQmoNlWP3NEzyhNhOtfKwLXSXKo-.-f2VYVWnfzV2aND- Passcode: k=4C6+Nu
https://byu.zoom.us/rec/share/st6f7PU56qdROPAFBCpNrlYazsE2vN1kMbBOYKtf8CiGe9p9_cCgEoZ8yPKdqowm.45TyIdvINlsWLl-u Passcode: T%Z1R97v