kornia/loftr

Pretrained weights for LoFTR (Detector-Free Local Feature Matching with Transformers), used by kornia.feature.LoFTR.

LoFTR builds dense coarse-to-fine correspondences directly from feature maps using linear attention, without any keypoint detection step. Trained on outdoor (MegaDepth) and indoor (ScanNet) datasets. CVPR 2021.

Original repo: zju3dv/LoFTR

Weights

File	Scene type
loftr_outdoor.ckpt	Outdoor (MegaDepth)
loftr_indoor_ds_new.ckpt	Indoor (ScanNet, updated)
loftr_indoor.ckpt	Indoor (ScanNet)

Citation

@inproceedings{LoFTR2021,
    title     = {{LoFTR}: Detector-Free Local Feature Matching with Transformers},
    author    = {Sun, Jiaming and Shen, Zehong and Wang, Yuang
                 and Bao, Hujun and Zhou, Xiaowei},
    booktitle = {CVPR},
    year      = {2021}
}