That's cool, this is connectivity clustering I was talking about: "Equipped with this scene graph, we can then retraverse the frames and assign the same label to each surface in the segmentation map that belongs to the same connected component in the scene graph.This allows distinct surface components to be identified as part of the same object across splits and joins over time (Fig. 6A,Row 5)."
It forms graphs, objects, surfaces, so it's still learning, but lateral. They should say that's not training, which is vertical. Their segmentation criteria is match of contiguous brightness and lateral shift vectors, and I could add other types of match to that. Segmentation is just a top-down perspective on connectivity clustering. I don't think any of that is terribly novel, just classical computer vision that people forgot after shift to neural nets. Because NNs are far easier to scale in complexity, from coding and parallelization perspective. That's the biggest problem with consistent connectivity clustering, it's whole lot harder to design and to utilize brute force than centroid-based clustering. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T943406c9c46e5cad-M1dc8952bae6182634f3cbe41 Delivery options: https://agi.topicbox.com/groups/agi/subscription
