Separate extraction and matching

[sam-s22]

I tried your model and it looks impressive. I’m curious to know if it's possible to extract features first and then perform matching. This approach could help me reduce the processing time for my specific application.

[chicleee]

Pre-extracting and storing features may not be suitable for semi-dense matchers, but it is theoretically feasible, though I haven't tried this yet.

In the current paradigm, only backbone features can be pre-stored (multi-scale integration requires f32/f16/f8 layers per image).

Feature extraction only accounts for about 1/4 of the time, so it may not be cost-effective depending on the scenario.

[sam-s22]

Thank you.

…

On Thu, Jul 3, 2025, 6:27 AM Xi Li ***@***.***> wrote: *chicleee* left a comment (chicleee/EDM#4) <#4 (comment)> Pre-extracting and storing features may not be suitable for semi-dense matchers, but it is theoretically feasible, though I haven't tried this yet. In the current paradigm, only backbone features can be pre-stored (multi-scale integration requires f32/f16/f8 layers per image). Feature extraction only accounts for about 1/4 of the time, so it may not be cost-effective depending on the scenario. — Reply to this email directly, view it on GitHub <#4 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BPEBYV4S4TEMNV5WSQD3S6L3GSPI7AVCNFSM6AAAAACAVGDV32VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTAMZQGUYDIMRYHA> . You are receiving this because you authored the thread.Message ID: ***@***.***>