DINOcular is a new approach to vision models that finally lets transformers natively “see” in 3-D. Instead of ignoring depth, it fuses RGB and depth in a single self-supervised backbone—injecting 3-D geometry into attention and patch embeddings.
The results: On 3-D
