5,921 human-validated object crops from children's egocentric video, across
72 CDI noun categories. Each point is one crop, laid out by t-SNE of its CLIP
representation at a chosen transformer block. Drag the layer slider to watch the
space reorganize from early visual features to the final semantic embedding used in the
CCN 2026 abstract. Scroll to zoom, drag to pan, double-click to reset.