Undergraduate student research project

Teaching a computer to watch mice the way a trained observer does

SqueakLab is a student study project exploring whether the social behaviour of laboratory mice can be scored automatically from ordinary top-view video — and, just as importantly, how wrong such a system is when it is wrong.

This is a study project, not a product. It is coursework-scale research carried out by an undergraduate student, published here so the work can be read and critiqued. Nothing on this site is a validated scientific instrument, it is not for clinical or diagnostic use, and it is an independent project — not an official project of, nor endorsed by, Brown University or any hospital or laboratory named in it.

The question

Behavioural studies of mice are scored by hand. A researcher watches video and marks every time two animals sniff, chase, groom, or fight. It is slow, and two trained people watching the same recording will not fully agree with each other.

The project asks a narrow version of a big question: given two visually identical dark mice in a standard cage, filmed from directly above, how much of that scoring can be done automatically — and can the remaining errors be made visible instead of hidden?

What has been built

Tracking

A silhouette tracker that follows two animals, and flags the moments it is unsure which mouse is which rather than guessing silently.

Scoring

Social and individual behaviours derived from the tracked geometry — distance, approach speed, relative orientation.

Human review

A workspace for correcting the machine's output, because the corrections are the most valuable data the project produces.

The honest part

Two mice of the same colour swap identities when they huddle. Rearing is nearly invisible from directly above. Fighting may be well under 1% of frames, so a model that never predicts it still looks accurate on paper. These are the interesting problems, and the project is written up around them rather than around a headline accuracy number.

Any figures shown in the workspace demonstrations are synthetic unless explicitly labelled otherwise.

The workspace

The analysis tool itself runs behind a login — it holds research video and is not public.