PyTorch implementation of "ResMatch: Residual Attention Learning for Feature Matching", by Yuxin Deng and Jiayi Ma.
08-2023 We upload new pretrained models to match new features, including DISK(NIPS2020), ALIKED(TMM2022,TIM2023), AWDesc(TPAMI2023). Performance on ScanNet is below:
| 5 | 10 | 15 | 20 | 5 | 10 | 15 | 20 | MS | P | |
|---|---|---|---|---|---|---|---|---|---|---|
| DISK | 32.0 | 42.0 | 48.8 | 53.8 | 13.4 | 28.5 | 38.2 | 45.1 | 14.22 | 48.02 |
| ALIKED | 37.1 | 47.8 | 54.5 | 59.1 | 16.4 | 33.1 | 43.4 | 50.2 | 14.51 | 48.80 |
| AWDesc | 42.9 | 54.3 | 60.8 | 65.4 | 18.7 | 37.7 | 48.6 | 55.6 | 11.97 | 48.19 |
07-2023 We upload a pre-release version composite of models, pre-trained weights. It might be enough to repoduce the results in the codebase of SGMNet. It will be a long time for the release of full codes since the paper is under review.