search

dcharatan / flowmap

[3DV 2025] Code for "FlowMap: High-Quality Camera Poses, Intrinsics, and Depth via Gradient Descent" by Cameron Smith*, David Charatan*, Ayush Tewari, and Vincent Sitzmann
https://cameronosmith.github.io/flowmap/
MIT License
893 stars 87 forks source link

where is inference code? #5

Closed zoldaten closed 6 months ago

zoldaten commented 6 months ago

i tried python3 -m flowmap.overfit dataset=images dataset.images.root=path/to/folder/with/images with Lighthouse image dataset but it takes huge time to got final result even on A100 Tesla. may be i started training not inference ? how to check it out ?

dcharatan commented 6 months ago

On an A100, the 150-frame Tanks and Temples sequences are expected to take ~22 minutes to run (see the last table in the supplemental). The script python3 -m flowmap.overfit is the correct one to run.

You might be able to speed up optimization by using the low_memory configuration option (see #4) or by playing around with mixed precision/FP16 settings. Also note that the lighthouse scene is one where FlowMap fails. See the code snippet below for notes on which scenes COLMAP and FlowMap fail on:

https://github.com/dcharatan/flowmap/blob/a6f854fb9376fdaf9c40b2a0cf81a2378424fd84/paper/common.py#L97-L155

zoldaten commented 6 months ago

ok, i tried Family dataset: python3 -m flowmap.overfit dataset=images dataset.images.root=Family i took ~ 20 min. but got strange result when open points3D file: 2024-04-27_09h41_07

images in dataset looks like -

00001

may be i do something wrong or the result should look like that ?