Edges in graph - Githubissues

schillij95 / ThaumatoAnakalyptor

Automatic Scroll Segmentation Pipline for CT Scans of Herculaneum papyri

MIT License

75 stars 13 forks source link

Edges in graph #26

Open giorgioangel opened 6 days ago

giorgioangel commented 6 days ago

https://github.com/schillij95/ThaumatoAnakalyptor/blob/613a8b1fb680509bbad51a57ad0ce63db65a56d3/ThaumatoAnakalyptor/instances_to_graph.py#L649

Given a patch p in a block A and an overlapping subvolume B, are we adding only in the graph the edge between p and the patch q in B for which the certainty is maximal? Isn't this too restrictive, especially if the estimate of the weight is not accurate?

schillij95 commented 6 days ago

Good point. It might be better to take the top n or so!

How would the ideal graph given our instances look like?

Some ideas:

Same Block edges
Other Block edges
Good certainty estimation function: What properties? Use distance from umbilicus? use 2.5D surface alignment? ...
Discard unlikely edges/nodes to make the problem smaller. (scalability)
...

giorgioangel commented 6 days ago

Top-n seems a good idea!

For the weight function we can use:

Sinkhorn divergence (approx of Wasserstein distance), this gives a similarity in shape
Cosine similarity between eigenvectors of PCA, this gives a proximity in orientation
Relative distance from umbelicus
Relative difference in error of fitting 2.5 surface
... Others?

We need a weight function that imposes the partial ordering we want and takes into account the highest number possible of "edge-cases"