open-forest-observatory / geograypher

Multiview Semantic Reasoning with Geospatial Data
BSD 3-Clause "New" or "Revised" License
10 stars 4 forks source link

Cache expensive results such as pixel-to-face correspondences #46

Closed russelldj closed 1 month ago

russelldj commented 7 months ago

It's very likely that we'll process the same dataset multiple times. It would be nice to cache some results so future runs can be accelerated. Since computing the pixel-to-face correspondences is expensive, and these results are the same for a given mesh+camera pair, it would be nice to save it to disk after it's computed. This could be done by writing a folder of .npy files with a parallel structure to the images, or using a more generic caching library, like ubelt.Cacher.

russelldj commented 2 months ago

I've been running experiments where I have to compute pix2face many times for a given mesh and camera set. For example, I'm running label rendering once, and then aggregating three separate sets of predictions. These four operations use the same set of pix2face correspondences.

asidhu0 commented 1 month ago

Resolved with #115