Research note

Reconstructing handwriting: define the target first

Are we trying to reproduce the ink path, recognise text, or both?

Vishnu KosuriJun 20262 min read

Revised 13 September 2026 · Background and proposed evaluation; no new Vahini benchmark reported

Sensor recording → Tip trajectory → Readable text. Explanatory diagram, not experimental data.

The question

Are we trying to reproduce the ink path, recognise text, or both?

Trajectory reconstruction predicts where the tip moved. Handwriting recognition predicts a sequence of characters. They can be connected, but the outputs and failure cases are different. A recogniser may output the right word despite a distorted path because language context fills the gap. Background: Fadeeva et al. (2024): Representing Online Handwriting for Recognition in Large Vision-Language Models; scikit-learn: data leakage and evaluation pitfalls.

A useful experiment therefore keeps a reference trajectory and a reference transcription when both outputs matter. It also labels pen-up motion. Training on clean tablet traces does not establish performance on inertial recordings collected in a different device and environment.

Worked example

A right word can hide a wrong path

For the word “cat”, a model outputting “car” has one substitution in three reference characters: character error rate is 1/3. That says nothing about whether the reconstructed t crossed at the correct location.

How to evaluate it

01
Define trajectory, contact and transcription outputs before training.
02
Keep test writers separate from training writers.
03
Report geometric error and character error separately, with examples of failures.