Add tanks and temples dataset loader #653
Conversation
…_barn/Barn_COLMAP_SfM.log
|
|
||
| retriever: | ||
| _target_: gtsfm.retriever.sequential_retriever.SequentialRetriever | ||
| max_frame_lookahead: 2 |
gtsfm/frontend/correspondence_generator/synthetic_correspondence_generator.py
Outdated
Show resolved
Hide resolved
| open3d_mesh_path = tempfile.NamedTemporaryFile(suffix='.obj').name | ||
| open3d.io.write_triangle_mesh(filename=open3d_mesh_path, mesh=mesh) | ||
|
|
||
| loader_future = client.scatter(loader, broadcast=False) |
There was a problem hiding this comment.
We could just avoid futures for this file I guess?
There was a problem hiding this comment.
Hmm wouldn't we lose all parallelization benefits then?
There was a problem hiding this comment.
Yes, but since we are just loading from disk, we would not lose much in terms of time I think
There was a problem hiding this comment.
I am wondering if we can just generate the matches apriori and use something like a cache to just load the matches from the disk.
There was a problem hiding this comment.
I was also thinking this would be the best approach, as this is more for debugging and will likely only be used by us.
There was a problem hiding this comment.
I see -- if we wish to use the Tanks and Temples loader to get GT poses to measure pose error, what's the difference between computing matches offline vs. online?
There was a problem hiding this comment.
Because the synthetic matching frontend is only used for debugging, and I'm not sure we should merge it into the main repo
There was a problem hiding this comment.
@ayushbaid @travisdriver wanted to revisit this -- I would prefer to keep as-is and make this code self-contained so that we only need to run one command, instead of having to make two new scripts (one to generate all the correspondences beforehand, and then making a new one to accept saved correspondences)
| min_score: 0.2 | ||
| max_frame_lookahead: 3 | ||
|
|
||
| # retriever: |
There was a problem hiding this comment.
Removed, thanks.
| _target_: gtsfm.frontend.correspondence_generator.synthetic_correspondence_generator.SyntheticCorrespondenceGenerator | ||
| #dataset_root: /Users/johnlambert/Downloads/Tanks_and_Temples_Barn_410 | ||
| #dataset_root: /usr/local/gtsfm-data/Tanks_and_Temples_Barn_410 | ||
| dataset_root: /home/runner/work/gtsfm/gtsfm/Tanks_and_Temples_Barn_410 # Path for CI. |
There was a problem hiding this comment.
I'd prefer this being loaded from the dataset root passed from the terminal, but its not an easy fix. So we should be fine with it.
| open3d_mesh_path = tempfile.NamedTemporaryFile(suffix='.obj').name | ||
| open3d.io.write_triangle_mesh(filename=open3d_mesh_path, mesh=mesh) | ||
|
|
||
| loader_future = client.scatter(loader, broadcast=False) |
There was a problem hiding this comment.
Yes, but since we are just loading from disk, we would not lose much in terms of time I think
Add image and mesh loader for scenes from the Tanks & Temple dataset.