search

Parskatt / RoMa

[CVPR 2024] RoMa: Robust Dense Feature Matching; RoMa is the robust dense feature matcher capable of estimating pixel-dense warps and reliable certainties for almost any image pair.
https://parskatt.github.io/RoMa/
MIT License
598 stars 50 forks source link

Use coco_20k data? #72

Open longzeyilang opened 2 months ago

longzeyilang commented 2 months ago

Hi, what should I trianing RoMa with coco_20k data augmentation? like https://github.com/verlab/accelerated_features/blob/main/modules/dataset/augmentation.py https://github.com/verlab/accelerated_features/blob/main/modules/dataset/augmentation.py how should I revise?

longzeyilang commented 2 months ago

I do not want use megadepth and scannet data, just like coco_20k data?

Parskatt commented 2 months ago

Basically you need to update the dataloader and loss. I used to have coco in DKM training. Wouldnt really recommend it though.

longzeyilang commented 2 months ago

why?

Parskatt commented 2 months ago

Because if you train dense matchers on homography they only do well on homography.

longzeyilang commented 2 months ago

ok , I want to train my own dataset, about 128*128 image size

longzeyilang commented 2 months ago

According to you, I want use scannet data to crop 128*128 image size to train. and detection in my own dataset. is it correct?