Slow inference speed

[manojs8473]

On my system with RTX 3080 8GB and Ryzen 9 6900hx, I get around 4 FPS for feature detection(input images were at 1280x720 resolution). Is there any way to increase the inference speed ?

Btw amazing work! I found DISK to be far more robust than superpoint+superglue, SoSnet in terms of matching at large change in rotation as well as illumination .Only bottleneck being the inference time. It's far too slow to be used on a real time pipeline(in my case its 25 FPS)

[jatentaki]

Have you tried running detection in half precision? DISK is not blazing fast at inference time because of convolutions at full image resolution. You could try further quantizing it or distilling into some more runtime-optimized backbone.

[jatentaki]

Also, as a note, unless you strongly subsample (threshold) the feature detections, at 1280x720 you will have so many detections that the matching step will become a bottleneck as well, possibly more significant than the detection... And if you were to subsample them strongly, maybe you can just go with smaller images in the first place?