Course project · Image classification

When is an image label clear?

A scarf-classification exercise that made dataset definitions and visual edge cases concrete.

The problem

The task was to classify whether a person in an image was wearing a scarf. A simple question becomes less simple when visually similar clothing appears in the examples.

The approach

The analysis focused on consistent labelling, including examples of visually similar garments such as a hijab. The point was to define the visual category consistently, not to infer anything about a person’s identity, beliefs or background.

A decision tree distinguishes clear matches, non-matches and edge cases when assigning a clothing label.Enlarge
Consistent image labels
Text description of the infographic

Each image is compared with the category definition. Clear matches and non-matches can be labelled consistently. Visually similar clothing is an edge case that calls for reviewing the definition. This concerns the garment, not the person’s identity or beliefs.

The challenge

Different label interpretations can make a dataset internally inconsistent even before model training begins.

Outcome & value

A concrete lesson in dataset design and classification ambiguity. The original notebook can be added when available; no model performance is presented without it.

Lessons learned

Document inclusion and exclusion criteria and review difficult examples before extending the training set.

Something to work on?

Let’s talk.

About data, AI, horticulture or a product you want to build.

Get in touch