I found the matching speed is slow ( use superpoint+lightglue )

3DOM-FBK / deep-image-matching

Multiview matching with deep-learning and hand-crafted local features for COLMAP and other SfM software. Supports high-resolution formats and images with rotations. Both CLI and GUI are supported.

https://3dom-fbk.github.io/deep-image-matching/

BSD 3-Clause "New" or "Revised" License

297 stars 32 forks source link

I found the matching speed is slow ( use superpoint+lightglue ) #10

Closed 16Huzeyu closed 8 months ago

16Huzeyu commented 8 months ago

First of all, I would like to thank you for your work. However, when I was using it to process 1k images of 4k resolution, I noticed that the matching speed was very slow, around 3 pairs/s. I saw that you have mentioned speed improvements in your todo, and if there is any progress, I will follow up promptly.

lcmrl commented 8 months ago

Hi, are you using a brute-force approach? The best thing to do is to pre-process your images to match only the overlapping ones,. You have different options: 1. --strategy matching_lowres preprocesses everything at low resolution and then matches at high res only the overlapping images 2. --strategy retrieval -r netvlad (or other global descriptors) computes a global descriptor for each image and compare all image pairs selecting only the best pairs to be processing at high res. 3. if for any reason you know the pairs to be matched you can use --strategy custom_pairs and pass a txt file with the pairs custom_pairs. In general to be a bit faster in matching you can change the config.py, see the options of "superpoint+lightglue_tiling" and change "quality": Quality.MEDIUM to "quality": Quality.HIGH. Probably you need to download/pull the updated version of the repository.

franioli commented 8 months ago

Hi, in addition to Luca's answer, we know there are some speed issues and we are working to find and reduce bottlenecks. Can you give us some more information about the configuration you are using? This might also be helpful for debugging.

By "configuration" I mean

Dimensions of your image, if you are using a tiling approach, and in this case the tiling grid and tile selection method (GRID, PRESELECTION) and tile overlap.
the matching strategy (brute-force, sequential, retrieval)
the superpoint and superglue configuration (if you are using one of the predefined configurations available in config.py or if you are using a custom one), in particular the maximum number of features to extract (per image or per tile).

Finally, may I ask you how you have found the code usability so far?

16Huzeyu commented 8 months ago

Thank you for your answers! @franioli @lcmrl I indeed use --strategy custom_pairs. And have tried superpoint+lightglue_fast ,it is almost 8xfatser but I need to test the quality . I found the code just by search "colmap+deep" in github beacause I have faced the problem how to imporve colmap with some new methods of feature extraction and matching.