Sometimes the model will not work as planned. Here are some things you can do to troubleshoot the process.
Crop differently. You can try different crops on the website, but you can also crop it on your own computer, especially if the image is slightly tilted, which means that a rectangle will not capture all of the text. When in doubt, try for less text, as long as you are capturing a full sentence.
Separate out different hands. The model can often manage several hands in the same image, but if they are different enough, or of different sizes (see example), then it is best to run the two separately by cropping them into two separate images.
Click here for an example of when two hands are different enough to sabotage the model.▶
Redraw lines. The lines should extend to the very end of the text. Sometimes the line segmenter will miss the end or beginning of the line, so check this step carefully. You might also need to add lines for interlinear additions, which will be transcribed as a separate line (rather than interpolated into the line below).
Adjust image. If the image is very dark or very light or a strange color (like green), the model will have trouble discerning the text. You can try changing the contrast on your own computer before uploading the image. Older images especially might need this adjustment.
