Data Collection

Where the photos came from

Handpicked, cropped, labelled

There is no large public collection of strabismus photos sorted by type, so we built one. We searched open-source image collections — Kaggle, GitHub and other public repositories — and handpicked photos where both eyes were clearly visible.

Each photo was cropped to the strip around the eyes and labelled with its class in CVAT, an open-source annotation tool. Shubh Garg and Dr. Debabrata Ghosh collected the photos; Madhav Arora and Bhumit Gupta labelled them, working through the set together, turn by turn.

The crop matters. The models only ever see the eyes, so they have to learn from how the eyes are aligned, not from faces, hair or backgrounds.

The dataset

517 photos in five classes

The classes are close to balanced: each has between 100 and 110 photos. Normal has slightly more, so during training every class was weighted equally, and no class could dominate simply by being bigger.