Hi everyone!
I’ve been working on a fast implementation of some newer local image feature detection/extraction methods for the past year. It’s somewhere between research code and a barely-usable library, GPLv3 licensed. This isn’t self promotion, but I really don’t have a use for it so I’m looking around if anyone maybe does (it’s not fun anymore just for the sake of it).
If someone says that ODM could use (or even need) this thing, it would keep me motivated to work on it, otherwise it will probably just rot away in its current state.
Cool project! It’s difficult to say whether it could be useful, without claims on quality compared to the current methods (DSP SIFT and/or SIFT). Maybe you could try to tie it into ODM and see if you get good results as a next step.
Welll for the task I was working on (very weird images tbh) it performed better than OpenCV SIFT, though not massively so (RANSAC/MAGSAC doesn’t need many falso positives to get confused). But I can’t publish that stuff, so I’d rather not make any prominent claims in writing haha.
The multi kernel descriptors by themselves are a real jump in robustness that’s uncontroversial, but the blob keypoint detector (same concept as in SIFT) is still a limiting factor. It just misses a lot of distinctive areas that you can only really find with big deep learning models.
So it could be useful in theory, that’s good to hear
Echoing Piero: it’s always hard to know what will be useful before it’s tested through the whole stack, but for sure: faster matching is a useful potential addition! And having a wider user base is a huge motivator I feel, so I definitely understand the desire to gauge interest.
