search

zju3dv / pats

Code for "PATS: Patch Area Transportation with Subdivision for Local Feature Matching", CVPR 2023
Other
90 stars 7 forks source link

Code release #2

Closed TuanTNG closed 1 year ago

TuanTNG commented 1 year ago

Hi, Thank you so much for your awesome work. When will you release your code?

xuanlanxingkongxia commented 1 year ago

We will release our code at the end of June. Thanks for your following.

TuanTNG commented 1 year ago

Hi @xuanlanxingkongxia, Thank you for your reply.

By the way, I have a question related to inference speed. May you help to compare the inference speed of your model with LoFTR?

xuanlanxingkongxia commented 1 year ago

It takes about 1 second for pats to match a pair of image, which is 10 times of LoFTR, while it is much faster than COTR and comparable with PDCNet+

TuanTNG commented 1 year ago

Thank you,

I got it. Looking forward to seeing your code.

xuanlanxingkongxia commented 1 year ago

In detail for our time and space efficiency: PATS (900ms) is slower than LoFTR (64ms) but comparable with PDC-Net+ (850ms) when processing 640×480 images. Besides, PATS (5.0GB) occupies more memory than LoFTR (3.6GB) but less memory than PDC-Net+(5.6GB). However, thanks to the scale-adaptive subdivision, for 1600×1200 images, PATS (8.3GB) consumes less memory than LoFTR (22GB) and PDC-Net+(23GB).

xuanlanxingkongxia commented 1 year ago

All of the tests are conducted on RTX 3090

Master-cai commented 11 months ago

@xuanlanxingkongxia hello, any update about the code release?

eugenelyj commented 10 months ago

Sorry for the delay. We have released the evaluation code and the pre-trained model. We are still working on the release of the training script, which is expected to appear at the end of this year.