[WIP] Bring some more methods from upstream and other minor fixes. #86
Conversation
| self.name2key[image.name]: image.image_id | ||
| for image in self.reconstruction.images.values() | ||
| } | ||
| # We cache the 3D points for mapping images to avoid parsing them from |
There was a problem hiding this comment.
Actually in #62 (review pending) I already had 100x speedup in pose estimation by only caching the image-to-point3D ID mapping, would this be sufficient? fetching the xyz seems already pretty fast and caching it is more expensive memory-wise. The locking adds quite a lot of complexity, no?
There was a problem hiding this comment.
I'll take a look there - maybe that's sufficient since I agree the locks complicate things.
To be honest, the fastest I managed to get our pose estimation was with single-threaded + caching. I didn't investigate too much where it's coming from though.
| 'cosplace': { | ||
| 'name': 'cosplace', | ||
| 'hloc': { | ||
| 'model': {'name': 'cosplace'}, | ||
| 'preprocessing': {'resize_max': 640}, | ||
| } | ||
| }, |
There was a problem hiding this comment.
cvg/Hierarchical-Localization@d0e8494 removes Cos/EigenPlaces but adds MegaLoc, which is much more robust (and has results on LaMAR).
There was a problem hiding this comment.
Thanks for the info, I'll go ahead and replace with MegaLoc and will also try that one on a currently private data split (to be released) to see where it stands (SALAD seems kinda good 😄 )
No description provided.