The output of the transformation layers are a collection of embeddings in the latent concept space. Those can be fed into an addition model to extract semantic

The output of the transformation layers are a collection of embeddings in the latent concept space. Those can be fed into an addition model to extract semantic segments, bounding boxes etc. IIUC this is how dinov3 works.